- Removed the VideoInfo field from the Chunk struct in chunk.go, streamlining the data model.
- This change reflects a shift in focus away from video information storage within the Chunk type.
`cached_tokens` is reported by every OpenAI-compatible provider that
supports prompt caching, but how it becomes non-zero differs by mode:
- Implicit caching (OpenAI, Azure OpenAI, DeepSeek, …) populates the
field automatically whenever a prompt prefix matches a previous
request within the provider's cache TTL. No client-side opt-in.
- Explicit caching (Qwen on Aliyun, Anthropic Claude, …) only
populates the field after the caller attaches `cache_control:
{"type": "ephemeral"}` to the relevant message / content block.
Until that opt-in is applied upstream of the request, the field
stays zero even when the prefix is otherwise byte-stable.
Without this distinction documented, the previous commit reads as if
`TokenUsage.CachedTokens` will show non-zero values for Qwen / Claude
once this PR lands — which is not the case. The plumbing here is a
prerequisite (stable prefix via sorted tools) and a meter (visibility
of the field), but the explicit-cache opt-in itself is out of scope
and lives elsewhere.
Document this on `TokenUsage.CachedTokens` and the `cachedTokens`
helper so callers do not mistake observability for activation.
OpenAI-compatible providers (Qwen, DeepSeek, OpenAI, Azure, etc.) report
prompt-cache hits in `usage.prompt_tokens_details.cached_tokens`. This
value was being read by go-openai but dropped at the WeKnora boundary,
so there was no way to tell whether prompt caching was actually working.
This change plumbs the field end-to-end:
- `types.TokenUsage.CachedTokens` (json:"cached_tokens,omitempty") —
zero-values are omitted so payloads stay quiet for providers that
never report cache hits.
- `cachedTokens` helper in remote_api.go guards against nil
`PromptTokensDetails` (Ollama and older OpenAI-compat backends omit
the details block entirely).
- All three response-parsing paths populate it:
* `parseCompletionResponse` (non-streaming)
* `processStream` (SDK streaming)
* `processRawHTTPStream` (raw-HTTP streaming, used when callers
need to inject custom fields like `cache_control`)
- The five `[LLM Usage]` log lines now print `cached_tokens=%d` so
cache hit rate is visible in `journalctl` / log tail without going
through metrics.
Together with the deterministic tool ordering from the previous commit,
this makes Qwen explicit caching observable: a warmed prefix should
show `cached_tokens` ≈ system + tools token count (typically several
thousand) on subsequent requests within the 5-minute TTL.
Tests:
- `TestCachedTokensHelper` — nil safety + round-trip
- `TestParseCompletionResponse_CachedTokens` — populated + missing
details paths through `parseCompletionResponse`
- `TestTokenUsage_CachedTokensJSONOmitempty` — zero is omitted,
non-zero is emitted
Lays type-system groundwork for the upcoming OpenSearch k-NN driver
(Phase 3 PR 2, see Tencent/WeKnora#1440), with strict feature-gate:
this PR ships only inert constants, schema extensions, and an
unreachable normalizer case. No path activates an OpenSearch
VectorStore yet — creation continues to fail with "not a valid engine
type" until the activation switch lands in Phase 3 PR 3.
Changes:
- OpenSearchRetrieverEngineType constant ("opensearch") in
internal/types/retriever.go. Not added to validEngineTypes /
GetVectorStoreTypes / retrieverEngineMapping yet (gated).
- ConnectionConfig.InsecureSkipVerify (bool, default false) in
internal/types/vectorstore.go, placed inside the // Common section
because it is a cross-driver TLS option. Distinct from the
Qdrant-specific UseTLS, which enables TLS on gRPC — InsecureSkipVerify
only controls verification of an already-TLS HTTPS connection.
AES-GCM Value/Scan round-trips the field as plaintext (it travels
alongside the encrypted Password / APIKey but is not sensitive itself).
- VectorStoreFieldInfo gains four optional fields: Immutable (bool),
Min/Max (*float64), Enum ([]string). All omitempty so existing UI
schema entries serialize identically. The fields will drive the UI
in Phase 3 PR 3 (read-only on Edit + range/enum constraints).
- Six new AuditAction constants under vector_store.* and opensearch.*
namespaces. Definitions only — emission lands in Phase 3 PR 3.
- EngineAwareNormalizer.Normalize is restructured to group engines by
the effective score range observed by the normalizer (not the
theoretical raw cosine range):
Range [-1, 1] (raw cosine, mapped via (score + 1) / 2):
- Milvus (COSINE metric mode). Milvus docs explicitly state the
COSINE metric range is [-1, 1].
Range [0, 1] (passthrough — already on the target scale):
- Elasticsearch v8 / ElasticFaiss. The driver issues a
cosineSimilarity(...) script_score script, and Lucene rejects
negative final scores ("Final relevance scores from the
script_score query cannot be negative" — ES docs); the
effective range observed by the normalizer is therefore [0, 1]
for IR-normalized embeddings. ES was previously grouped with
Milvus and over-corrected via (score + 1) / 2, which inflated
every ES result by 50% in mixed-engine RRF fusion.
- OpenSearch. The k-NN plugin's
SpaceType.COSINESIMIL.scoreTranslation maps the underlying
Lucene/Faiss distance (1 - cosine) to (1 + cosine) / 2 ∈ [0, 1]
before the score reaches us. Source:
github.com/opensearch-project/k-NN at
src/main/java/org/opensearch/knn/index/SpaceType.java
(COSINESIMIL enum, scoreTranslation method).
- Weaviate. The driver requests `certainty`, defined by Weaviate
as (2 - distance) / 2 = (1 + cosine) / 2, intrinsically [0, 1].
- Postgres pgvector / SQLite sqlite-vec / Qdrant /
TencentVectorDB / Doris. The driver computes (1 - cosine_distance)
or normalized inner_product, whose theoretical range is [-1, 1].
The IR-normalized positive-component unit vectors WeKnora
targets (BGE / OpenAI text-embedding-3 / Cohere /
sentence-transformers) keep the observed range in [0, 1];
negative-cosine embedding models would silently clamp to 0
downstream — explicitly documented as the IR-normalization
caveat in the struct godoc.
Dead enum references (ElasticFaiss, Infinity) are flagged in the
godoc with a pointer to internal/types/vectorstore.go's existing
"legacy/experimental, no standalone deployable instance" annotation.
Their case labels are kept for switch exhaustiveness.
Test coverage:
- OpenSearch constant wire value + collision check against existing
10 engines + gated invariant (NOT in validEngineTypes).
- ConnectionConfig.InsecureSkipVerify backward-compat (missing JSON
field deserializes as false) + round-trip + AES-GCM coexistence
with encrypted Password / APIKey.
- VectorStoreFieldInfo omitempty preservation + new-field serialization
+ *float64 pointer distinction (min=0 vs nil).
- AuditAction dot-namespace convention enforcement, prefix invariants
for vector_store.* and opensearch.*, no wire-string collisions, exact
wire-value pins for the six new constants.
- EngineAwareNormalizer:
- CosineRange retains its (score + 1) / 2 coverage but now only for
Milvus.
- UnitInterval now covers the full passthrough group (ES /
ElasticFaiss / OpenSearch / Weaviate / Postgres / SQLite / Qdrant /
Infinity / TencentVectorDB / Doris).
- New TestEngineAwareNormalizer_ElasticsearchCosinePassthrough is an
explicit regression guard for the score-range correction: cos=0.5
maps to 0.5 (not (0.5 + 1) / 2 = 0.75 as an earlier draft assumed).
- OpenSearch passthrough across (0 / 0.5 / 0.75 / 1) + engine drift
(1.0001) + defensive negative + ±Inf / NaN edges + keyword
passthrough + nil-ctx safety.
Backward compatibility is preserved at every layer:
- All new struct fields are omitempty / pointer-tagged so existing
rows and existing wire formats remain unchanged.
- Normalizer's new OpenSearch case is unreachable until the driver
lands. The ES regrouping changes the post-normalization value for
every ES vector search result (a correctness fix, not a feature) —
ES vector retrieval is currently the only production path affected.
- AuditAction constants emit no audit_log rows in this PR.
- engine_type=opensearch VectorStore creation still rejected (gated).
Backend
-------
ListKnowledgeBases now enriches each row with the resolved vector_store
metadata (name / source / engine_type / status) via a new
buildKBListResponse helper. Store views are batch-resolved once per
request through BatchResolveStoreView so an N-KB list costs one
vector-store service call rather than N — closing the N+1 limitation
flagged in #1372's known-limitations section. Cross-tenant shared KBs
continue to render via SharedStoreDisplay so the owning tenant's store
inventory cannot be correlated across rows; the underlying vector_store_id
UUID is stripped from those responses.
Resolver failures degrade gracefully: bound KBs render as unavailable
instead of breaking the list. Test coverage pins the env / bound /
shared distinction, the batch-call-count invariant, and the
graceful-failure path.
Frontend
--------
KB editor modal gains a new "Vector Store" section. Create mode shows a
dropdown that combines the system default (env store) and the tenant's
configured user stores, fetched once at mount via the existing
listVectorStores API. Edit mode shows the bound store read-only via a
new VectorStoreBadge component with an explicit immutability hint —
matching the backend's `<-:create` GORM tag and the service-layer
UpdateKnowledgeBaseRequest DTO that already omit the field.
KB list cards surface a small engine-type badge for own-tenant bound
KBs, and a warning badge when the bound store is unavailable. Env-bound
and shared KBs render no badge (visual noise control). KB detail header
shows the bound store via the same VectorStoreBadge component; shared
KBs fall through to the badge's internal "shared" branch with no name /
engine / id rendered.
The KB editor's create-time error handler translates the typed
ErrVectorStoreBindingInvalid (2200) and ErrVectorStoreUnavailable
(2201) into localized messages and jumps the user back to the
VectorStore section so they can pick a different store or fall back to
the system default.
The KB row type gains five optional fields (vector_store_id / name /
engine_type / source / status). i18n: 18 new keys added to en-US,
ko-KR, zh-CN; ru-RU receives English placeholders pending translation
(consistent with prior PRs in this locale).
Part of #993 (Phase 2: Per-KB VectorStore Binding).
Phase 2 roadmap item: PR 5 (KB binding UI + list-response enrichment).
Depends on #994, #1310, #1372, #1386 (all backend in the Phase 2 series).
Add support for configuring vLLM server URL when using vlm-http-client or hybrid-http-client backend in MinerU.
Changes:
- internal/types/tenant.go: Add MinerUVLMServerURL field to ParserEngineConfig
- internal/infrastructure/docparser/mineru_converter.go: Pass server_url to MinerU API when backend is vlm-http-client or hybrid-http-client
- frontend/src/api/system/index.ts: Add TypeScript type definition
- frontend/src/views/settings/ParserEngineSettings.vue: Add vLLM server URL input field
- frontend/src/i18n/locales/*.ts: Add translations (zh-CN, en-US, ko-KR)
Sessions now record the input-bar state used for the most recent QA
request (agent, model, KB scope, web search). The chat UI hydrates
those settings on session reopen so users see the same configuration
they used last time, instead of the global default.
The state is stored in the existing sessions.agent_config JSONB column
to avoid a new migration. Frontend snapshots the user's global defaults
on session enter and restores them on session leave, so opening an old
session does not pollute new-chat defaults.
Until now a fresh login (new device, expired refresh token, cleared
browser) always dropped the user into their home tenant, even if they
spend most of their time in a peer workspace. The session-local
X-Tenant-ID override in localStorage gave a "same browser sticky"
effect, but never crossed devices or new sessions.
This adds a small server-side preference, `users.preferences.
last_active_tenant_id`, persisted in the existing jsonb column (no
new migration), and threads it through:
* Backend
* `UserPreferences` gains `LastActiveTenantID *uint64` with sentinel
semantics (`*0` from the PATCH endpoint = clear preference).
* `resolveLoginTenantID` validates the stored id (tenant still
exists + active membership, or CanAccessAllTenants) and falls
back to home on any failure, best-effort clearing the stale
preference so subsequent logins don't pay for it again.
* `Login` and `LoginWithOIDC` resolve once and use the result for
both the JWT `tenant_id` claim and the returned `active_tenant`,
keeping the two in sync. `RefreshToken` rides through
`GenerateTokens` so refresh rotations also land the user back in
their preferred tenant instead of bouncing to home.
* `UpdateUserPreferences` learns to merge the new key.
* `PUT /auth/me/preferences` accepts the new field.
* Frontend
* `Login.vue` now writes the user's HOME tenant id into
`user.tenant_id` (matching the field's documented semantics) and
expresses any active-vs-home divergence via `setSelectedTenant`,
so `useHomeTenant` and the "current"/"home" badges stay correct
after the backend honours a remembered preference. `App.vue`'s
OIDC sync does the same reconciliation.
* `TenantSelector`, `UserMenu` and the post-tenant-create handlers
fire `persistLastActiveTenantPreference` after every successful
user-initiated switch (switching to home sends `0` to clear).
The call is raced against the existing reload-grace window so
most writes finish before the page tears down; lost writes are
recoverable on the next switch.
No new UI. Users will simply notice that, after re-logging in on
another device, they land back in the workspace they were last
using rather than always in their home tenant.
Note: `make docs` is unrelated-broken on `main` (audit_log.go
references `errors.AppError` which swag can't resolve), so the
Swagger artifacts under docs/ are intentionally not regenerated
in this PR. The handler code is the source of truth.
Refactored the pinning mechanism for knowledge bases to be user-specific, allowing any user with read access to pin knowledge bases independently. This change includes the introduction of a new `user_kb_pins` table to manage pin states, replacing the previous tenant-wide pinning system. Updated relevant API endpoints and service logic to support this new functionality, ensuring that the UI reflects the user's personal pin status.
Additionally, adjusted the knowledge base listing to sort by user-specific pin status, enhancing the user experience by prioritizing pinned items.
Refs: #1303
Added functionality for managing user preferences, including the ability to update preferences via a new API endpoint. The preferences are now stored server-side, allowing for synchronization across devices. Updated relevant components to handle user preferences, including the UserMenu and settings views. Enhanced internationalization support for error messages related to preference updates.
Introduced a new optional field, creator_name, in both CustomAgent and KnowledgeBase types to allow the front end to display the creator's name. This enhancement enables better differentiation between resources created by the current user and those created by other members of the same tenant. Updated relevant API handlers to populate this field during list operations, ensuring accurate representation in the UI. Additionally, modified the ResourceOriginBadge component to utilize the creator_name for improved context in resource listings.
Refs: #1303
Added functionality for managing pending invitations, including a polling mechanism to update the invitation count in real-time. Introduced a new dialog component for users to view and respond to their invitations without navigating away from their current context. Updated API endpoints for fetching and managing invitations, and enhanced the user interface with relevant internationalization support.
Refs: #1303
Introduced a new API client for managing user favorites, allowing users to list, add, and remove starred resources. Enhanced the sidebar to display favorites and recents, improving user navigation. Updated internationalization files to include relevant labels and tooltips for the new features across multiple languages.
Refs: #1303
Added a new configuration option to limit the number of tenants a non-superuser can create via self-service. Introduced a new error type for handling cases where users exceed this limit, returning a 429 status code. Updated the tenant creation and update handlers to enforce this limit and provide appropriate feedback to users. Additionally, refactored the tenant update request structure to ensure only mutable fields are allowed.
Refs: #1303
Updated the organization membership model to operate at the tenant level, aligning with Plan 3. This change introduces new interfaces and modifies existing ones to accommodate tenant-centric operations, including the `InviteMemberRequest` and `TenantInviteCandidate`. The search functionality for inviting members has been updated to reflect this shift, allowing for tenant-based searches instead of user-based. Additionally, internationalization files have been updated to ensure consistent messaging across languages regarding tenant membership and roles.
Refs: #1303
The audit_logs table grows monotonically until something purges it.
PR 6 explicitly listed retention as out of scope for v1; this picks
that up.
Adds an `audit.retention_days` config (default 90, 0 disables) and a
small background goroutine `AuditLogRetentionRunner` that fires once
~10 minutes after boot and then every 24h, calling
`AuditLogService.Purge` which DELETEs rows older than the cutoff in a
single statement. The dependency surface is intentionally minimal —
no robfig/cron, no asynq, no migration — because retention has no
wall-clock alignment requirement.
Wiring is the same shape as the existing data source scheduler:
container `Provide(NewAuditLogRetentionRunner)`, an `Invoke` that
calls `Start`, and a `ResourceCleaner.RegisterWithName` for graceful
shutdown.
Defaults preserve operator intent:
- `audit:` section omitted from YAML -> retention_days=90.
- explicit `audit.retention_days: 0` in YAML -> purge disabled.
- `WEKNORA_AUDIT_RETENTION_DAYS=N` env override (incl. N=0).
- `retention_days < 0` is rejected by ValidateConfig.
Tests cover the service Purge contract (no-op when disabled, cutoff
math, error propagation) and the runner lifecycle (Start no-op when
disabled, idempotent Start/Stop, ticker cadence, runOnce swallows
errors). 10 new tests, all green.
docs/rbac.md updated with the new YAML / env / behaviour.
Refs: #1303
Plan 3 (#1303) lifted Org membership from per-user to per-tenant, and
40fdb978 pinned the org's owner tenant in DB as organizations.owner_tenant_id.
The OrganizationSettingsModal members list, however, still decided
"which row is the owner" by comparing member.user_id against
orgInfo.owner_id. After Plan 3 the members list is keyed on tenant_id,
so the user-id check is structurally wrong: the representative user of
the owner tenant happens to be the owner today, but if that user is
moved to another tenant — or a different user of the owner tenant
becomes the representative — the modal would either lose the "(owner)"
tag, or re-tag a non-owner row, and would offer the role-select /
remove-button affordances on what is in fact the owner tenant row.
This change:
- Adds OwnerTenantID to OrganizationResponse and populates it in
toOrgResponse from the persisted column. IsOwner now compares the
caller's current tenant against owner_tenant_id (with a user-id
fallback for legacy rows where owner_tenant_id == 0).
- Adds owner_tenant_id to the frontend Organization type and a small
isOwnerMember(member) helper in OrganizationSettingsModal, which
tenant-matches by default and falls back to user-id only on
pre-000046 rows. The role <select>, role tag suffix, and the
delete button on member rows now all use the helper.
The "is-me" CSS class and the me-tag in the member name still match
on user_id — that one really is a per-user marker ("this row's user is
me, the signed-in user").
Plan 3's isOwnerTenant resolved "is this tenant the owner tenant?"
by looking up the owner user's CURRENT tenant via userRepo. That
breaks silently the moment the owner user is moved to another
tenant: the formerly-protected owning tenant becomes removable
(can be kicked from OTM, role-downgraded), and the new tenant the
owner moved to is treated as untouchable even though it isn't even
in OTM.
Promote owner-tenant to a first-class column on `organizations`:
- New migration 000046 adds owner_tenant_id BIGINT NOT NULL with
index. Backfill: (1) owner user's current tenant when known,
(2) earliest Admin tenant in OTM as fallback for orphans, then
(3) RAISE EXCEPTION on anything still NULL — refuse to silently
ship a bogus value. SQLite init schema mirrors the column.
- CreateOrganization now sets OwnerTenantID at create time so the
pinned value is correct for all new orgs.
- isOwnerTenant becomes a pure equality check on the persisted
column — no userRepo lookup, no error tuple. Fail-closed on a
zero column value (legacy data not yet backfilled or unit tests
bypassing the migration) so the worst case is "membership table
frozen for that org" rather than "owner tenant is removable",
matching the prior fail-closed-on-error semantics.
- RemoveTenantMember / UpdateTenantMemberRole simplify to a single
bool check.
Before this change, the `organization_members` table keyed on user_id —
joining an org as a Viewer in tenant T meant only that one user got
visibility on shared KBs/agents, while the rest of T (including its
Owner) saw nothing. The mismatch surfaced as confusing UX bugs:
"my colleague shared this with us, why can't I see it?"
This change replaces `organization_members` with
`organization_tenant_members`. Membership and Org-RBAC role now key on
(org_id, tenant_id); `representative_user_id` stays as a display-only
label of the user who first brought the tenant in.
Permission resolution gains a 3-D cap so tenant-RBAC stays honest under
cross-tenant sharing:
effective = min(share.Permission, tenant_org_role, tenant_role_cap)
where tenant_role_cap pins tenant Viewers to OrgRoleViewer regardless
of the org-level grant. A Viewer in their own tenant cannot edit a
shared KB even if the org granted Editor — preserves the
"Viewer = read-only everywhere" promise.
Migration 000045 backfills `organization_tenant_members` from the old
table by collapsing each (org, tenant) group to its max role, then
renames the old table to `organization_members_pre_plan3` for safe
rollback. Sqlite mirror (Lite mode) is fresh-bootstrap: the init
schema replaces the old table directly.
Routes: member-management URLs change from `:user_id` to `:tenant_id`.
Org create / join / leave now require Admin+ in the caller's tenant —
joining an Org affects everyone in T, so the decision can't sit with
a Viewer.
Frontend: OrganizationSettingsModal + store update to address members
by tenant_id.
The Tenant.ConversationConfig field was marked deprecated when
CustomAgent (builtin-quick-answer) replaced the per-tenant defaults
path, but the JSONB column, GET/PUT KV handlers, write helpers and
read fall-throughs all stayed wired. This had two concrete failure
modes:
- Input-field.vue silently PUT-ed per-user model picks back into
/tenants/kv/conversation-config, which requires Admin+ — Viewers
and Contributors got a 403 every time they switched chat models.
- The same write also overwrote the tenant-shared default for every
other user, turning a per-user UX into a per-tenant footgun.
Remove the field, the struct, the validate helper, the PUT case and
the two dead Summary helpers in session/helpers.go. KnowledgeSearch
loses its tenant-conversation fallback (one chunk-tool reader); the
global config + hardcoded defaults still cover it. session_agent_qa
already required agents to declare their own rerank_model_id, so no
behaviour change there.
Frontend follows:
- Input-field.vue: per-user 'last selected chat model' moves to
localStorage, no API call.
- AgentEditorModal: defaults for builtin-quick-answer creation now
come from prompt-templates (system/context/rewrite/fallback) and
retrieval-config (top_k / threshold), the two endpoints that
actually own those settings now.
- AgentSettings.vue: orphan settings page with no router/import
referencing it; deleted.
- api/system: getConversationConfig / updateConversationConfig and
the ConversationConfig interface go away.
Net: 137 insertions, 2761 deletions.
PR 1-5 of the multi-tenant RBAC series enforce the role matrix; what
they don't do is keep a durable record of who did what. Today the
RBAC middleware logs "[rbac] role insufficient ..." to stderr -
perishable, not searchable. Operators cannot answer "who removed
Alice from tenant T-foo on Tuesday" or "did we 403 a probing user
200 times last hour" after the fact.
This PR adds a generic per-tenant audit log that:
- Captures mutating RBAC actions (member add/remove/role-change/leave)
durably from the tenant_member service after the repo write succeeds.
- Captures enforcement rejections (RequireRole / RequireOwnershipOrRole
denies under EnableRBAC=true) via a gin-context-injected hook.
- Sliding-window dedup (1 minute) on (tenant, actor, action, path)
for denied events so a probing client cannot flood audit_logs.
- Default-on - no extra env flag. Schema is generic enough for future
KB / agent / datasource events to reuse via new action namespaces
without another migration.
Schema (migration 000044) mirrors wiki_log_entries' shape: a single
audit_logs table with action, target_type/id, request path/method,
outcome, and a JSONB details column. Three indexes: (tenant_id, id
DESC) for the per-tenant cursor feed, actor_user_id for "what did
Alice do", (tenant_id, action) for both the action filter and the
dedup lookup. Sqlite mirror added for Lite mode.
Wiring is nil-safe: the audit field on tenantMemberService and the
gin-context lookup in rbac.go both tolerate a nil service so existing
tests construct without a stub and dormant deployments degrade
gracefully if the provider is misconfigured.
The query API is GET /tenants/:id/audit-log, gated by Admin+ on top
of the existing PathTenantMatch (denial histories should not surface
to ordinary members). Cursor-paginated by descending id (sidesteps
the duplicate-timestamp tie-breaking that BIGSERIAL trivially solves).
Frontend: TenantMembers.vue gets a t-tabs wrapper with Members
(existing) and Audit log (new) tabs; the audit tab is gated to
Admin+ to mirror the server, lazy-loads on first open, paginates via
the next_cursor pattern, and renders coloured chips per
action/outcome so an operator can scan for anomalies. i18n keys
added in all four locales.
Tests: 14 new tests across service (6), middleware (4), and handler
(4). Service tests pin the dedup window behaviour (writes again
after expiry, dedup is per-actor-per-path, degrades on lookup
error). Middleware tests pin both reject sites' audit hook and the
"dormant mode does not fire" + "nil audit does not panic"
invariants. Handler tests pin the cursor envelope and that an empty
page returns next_cursor=0.
Out of scope: per-action ACL on the audit log, retention/TTL job,
KB/agent/datasource event wiring (schema supports it - just no
emitters yet), async writes, frontend filter UI.
PR 1–4 of the multi-tenant RBAC series gated KB-level and Agent-level
mutations behind a "creator OR Admin+" check via OwnedKBOrAdmin /
OwnedAgentOrAdmin. The layer underneath — knowledge entries, chunks,
wiki pages — was still gated only by Contributor, which is wider than
intended: any tenant Contributor could edit/delete documents in a
private KB they didn't own. PR 2 / wiki author left explicit follow-up
breadcrumbs at the three call sites in router.go acknowledging this gap.
This PR closes those breadcrumbs by reusing the existing scaffolding:
- KnowledgeService.GetOwningKBCreatorID walks knowledge_id -> kb_id ->
KB.CreatorID, tenant-scoped through GetKnowledgeByID. The chain lives
in the service layer so the handler-side lookups stay pure adapters.
- Three new CreatorLookup closures (KBCreatorLookupFromKnowledgeID,
KBCreatorLookupFromKnowledgeIDParam, KBCreatorLookupFromKBPath) are
added in handler/rbac_lookups.go alongside the existing
KBCreatorLookup / AgentCreatorLookup. Each translates the repository
sentinels into middleware.ErrResourceNotFound and re-checks tenant
scope as defence-in-depth, mirroring KBCreatorLookup exactly.
- router/rbac.go gains three guard factories — OwnedKnowledgeKBOrAdmin,
OwnedChunkKBOrAdmin, OwnedWikiKBOrAdmin — wired through the same
rbacGuards bundle.
- Mutating routes in RegisterKnowledgeRoutes / RegisterChunkRoutes /
RegisterWikiPageRoutes switch from g.Contributor() to the new
ownership guards. Cross-KB batch operations (POST /knowledge/move,
POST /knowledge/batch-delete, PUT /knowledge/tags) stay
Contributor-gated since there's no single owning KB to check
against. KB-scoped uploads (POST /knowledge-bases/:id/knowledge/...)
reuse the existing OwnedKBOrAdmin because the URL :id is the KB id
directly. The chunks.DELETE("/by-id/:id/questions") route addresses
chunks by their own id with no knowledge id in the URL, so it stays
Contributor-gated as a documented carve-out.
No schema changes: documents inside a shared KB are KB-owned content,
not personal uploads — that's how the product is positioned and matches
the original "per-KB-ownership granularity" wording from the PR 2
breadcrumbs. Legacy/built-in KBs with empty CreatorID surface as
tenant-owned (Admin+ only), same rule as built-in agents.
Tests:
- 9 new test cases in handler/rbac_lookups_test.go covering the three
new lookups: NotFoundMapsToSentinel (each repository sentinel),
CrossTenantIsHiddenAsNotFound, OwnerMatchReturnsCreatorID,
MissingTenantContext (fail-closed), and ChunkRouteShape (param-name
regression guard).
- 5 new test cases in service/knowledge_owner_chain_test.go covering
the chain helper: knowledge-not-found, kb-not-found, nil-kb-without-
err, happy path, and legacy KB with empty CreatorID.
Build, vet, and the targeted ./internal/middleware/...
./internal/handler/... ./internal/router/...
./internal/application/service test runs are clean. Pre-existing
TestCreateStore_* / TestOssEnsureBucket_CreateFails failures on main
are unrelated and reproduce on a clean checkout.
Security
- Reject URL :id when it does not match the caller's active tenant on
every /tenants/:id/members/* and /tenants/:id/leave endpoint. Without
this an Owner of tenant A could drive operations on tenant B simply
by changing the URL, because RequireRole resolves the caller's role
against the active tenant from context, not the URL param. The
cross-tenant superuser carve-out mirrors middleware/rbac.go
(CanAccessAllTenants + EnableCrossTenantAccess required).
- AddMember no longer records the synthetic "system-<tenantID>" caller
(X-API-Key auth path) as invited_by, matching the same filter
KB.CreatorID got in PR 2.
- AuthHandler now logs a loud one-shot error when constructed with a
nil/incomplete config so a wiring bug that silently disables the
invite_only gate is visible at boot.
Race fixes
- New repo helpers DemoteOwnerAtomically / RemoveOwnerAtomically wrap
the count-other-owners + write step in a transaction that holds
SELECT … FOR UPDATE locks on every other active Owner row. The
service routes Owner demotion / Owner removal through them, closing
the TOCTOU window where two concurrent demotes of two different
Owners could each observe count=2 and both commit, leaving the
tenant ownerless.
Performance
- ListMembers hydration now batches user lookups via a new
GetUsersByIDs interface method (one SELECT … WHERE id IN (…)
instead of N round-trips). Failure degrades to "no email/username"
rather than dropping rows.
Frontend
- canManage also fires for cross-tenant superusers
(authStore.canAccessAllTenants), mirroring the backend carve-out so
the UI controls match what the server actually allows.
Tests
- Cross-tenant URL guard regression cases for List / Add / Update /
Remove / Leave.
- Superuser bypass tests (flag on -> service reached, flag off -> 403).
- Batched lookup test ensures GetUsersByIDs is called exactly once and
GetUserByID is never used as a fallback.
- Synthetic caller invited_by filter regression test.
- Service-layer atomic last-Owner regression tests for both
UpdateRole and RemoveMember.
Wire the human-facing surface for the tenant-RBAC system that PR 1
introduced and PR 2 enforced: REST endpoints + UI for managing tenant
members, plus the long-pending invite_only registration gate.
Backend
- internal/handler/auth.go: Register now rejects with 403 when
auth.registration_mode=invite_only. PR 1 already exposes the mode via
/auth/config so the frontend hides the sign-up link; this is the
server-side enforcement for direct API calls.
- internal/handler/tenant_member.go: new TenantMemberHandler with
ListMembers / AddMember / UpdateMemberRole / RemoveMember. Maps
service-layer sentinels onto HTTP semantics (ErrLastOwner -> 409,
ErrMembershipAlreadyExists -> 409, ErrUserNotFound on AddMember -> 404,
ErrInvalidTenantRole -> 400). AddMember requires the invitee to
already have an account; SMTP-driven email invites are deferred.
- internal/types/tenant_member.go: new TenantMemberResponse projection
carrying the fields the management UI renders (email, username,
avatar, role, joined_at).
- internal/router/router.go: four routes under /tenants/:id/members
gated per the matrix (Viewer+ for list, Owner+ for every mutation).
- internal/container/container.go: dig provider for the new handler.
Tests
- handler/tenant_member_test.go: 12 table-driven cases covering happy
path, dangling-user tolerance, ErrLastOwner -> 409, ErrMembershipNotFound
-> 404, ErrMembershipAlreadyExists -> 409, unknown email -> 404,
invalid role rejected before user lookup, bad tenant id -> 400.
- handler/auth_register_invite_only_test.go: invite_only -> 403,
self_serve -> 201, nil Auth config falls through.
Frontend
- src/api/tenant/members.ts: typed client (TenantRole/TenantMember/...)
for the four endpoints.
- src/views/settings/TenantMembers.vue: members table with add dialog,
inline role select, confirm-remove modal. All mutation widgets gated
behind currentTenantRole === 'owner'; server still enforces.
- src/views/settings/Settings.vue: register the Members tab; show only
when currentTenantRole is loaded (avoids tab flash on cold nav).
- i18n in en-US/zh-CN/ko-KR/ru-RU: new tenantMember namespace covering
all UI strings and structured error messages (lastOwner / userNotFound
/ alreadyMember / invalidRole etc).
This completes the #1303 RBAC rollout. With cfg.Tenant.EnableRBAC=true
operators now have a full management surface; with EnableRBAC=false
the routes still exist but mutations log-and-pass per PR 2 contract.
Middleware
- RequireOwnershipOrRole now skips the creator lookup entirely when
EnableRBAC is off, eliminating a hidden per-request DB roundtrip in
the rollout window.
- New ErrResourceNotFound sentinel lets lookups signal "not visible to
this tenant"; the middleware passes the request through so handlers
can return a real 404 instead of a misleading 403.
- Transient lookup failures now surface as 503 (was 403) so monitoring
can tell denied-by-policy from cannot-verify-right-now.
- Cross-tenant superusers (CanAccessAllTenants) bypass tenant role gates,
restoring DELETE /tenants/:id and ResetAPIKey for org operators that
resolve to Admin in foreign tenants.
- Construction-time warning fires once when cfg is nil/incomplete, so a
wiring bug that disables enforcement globally is visible at boot.
Route coverage
- Add role guards to FAQ, KnowledgeTag, WikiPage, Skill, WebSearch,
WebSearchProvider, WeKnoraCloud, Session, Chat, and Message route
groups so the matrix actually matches the PR description.
Service layer
- KB and agent creators are no longer recorded as the synthetic
system-<tenantID> user used by the X-API-Key path; resources stay
tenant-owned. New types.IsSyntheticUserID helper centralises the
prefix check.
- KBCreatorLookup and AgentCreatorLookup re-check tenant scope so a
same-ID resource in another tenant cannot piggyback through the
ownership shortcut.
Docs
- CustomAgent.RunnableByViewer comment now matches enforcement: it only
blocks Viewer-driven agent-mode runs, not plain RAG/Wiki QA.
Tests
- New cases for the 503/404 paths, lookup-skip behaviour, cross-tenant
superuser bypass, the lookup helpers (not-found, cross-tenant,
built-in agent, missing tenant context), and IsSyntheticUserID.
POST /auth/switch-tenant mints an access token whose `tenant_id` claim
points to the target tenant, but the middleware previously discarded
that claim entirely:
user, err := userService.ValidateToken(ctx, token) // returned only *User
targetTenantID := user.TenantID // always the home tenant
The session therefore stayed scoped to the user's home tenant despite
the supposedly fresh token, making the new endpoint a no-op once
the request hit any tenant-scoped query.
This change widens the contract:
- UserService.ValidateToken now returns (user, activeTenantID, error)
where activeTenantID is the JWT's `tenant_id` claim, falling back to
user.TenantID when the claim is absent (backward compatibility with
pre-RBAC tokens still in the wild).
- middleware/auth.go uses activeTenantID as the default target before
applying any X-Tenant-ID header override, and recomputes
crossTenantSwitch from `activeTenantID != user.TenantID` so the
resolveTenantRole branches stay correct for switch-tenant sessions.
- The non-middleware /auth/validate handler in handler/auth.go absorbs
the new return value (it only cares about user identity).
Tests:
- internal/application/service/user_jwt_test.go pins the JWT->tenant-id
projection (`tenantIDFromClaims`) — claim present beats fallback,
zero/negative/string claims fall back, int64/uint64 are accepted for
test-built tokens.
Required for PR 2 multi-tenant integration tests, where a single user
needs to switch between Owner-in-A and Viewer-in-B sessions and observe
distinct authority on each side.
Tighten correctness, security and observability of the tenant RBAC
foundation introduced in 3c4cdd70 in response to internal review.
Bugfixes
- custom_agents.runnable_by_viewer: drop the GORM `default:true` tag.
GORM treats a plain bool's zero value as "unspecified" and silently
rewrites it to the column default on insert, which would prevent
admins from ever creating an agent with viewer-runtime disabled. The
column-level DEFAULT TRUE on the database side still backstops rows
inserted by raw SQL / migrations.
- TenantMemberRepository.HasAnyMembers: replace `Limit(1).Count(&n)`
with `Take(&probe)`. GORM's Count ignores Limit and scans the whole
table, which matters because this runs on the auth middleware's
hot path for users without a cached membership.
- service/user.LoginWithOIDC: include User, Tenant and Memberships in
the OIDCCallbackResponse so the OIDC flow returns the same shape as
password login (previously User and friends were left nil).
Security
- middleware/auth.resolveTenantRole: restrict orphan-tenant
auto-promotion to the caller's home tenant. Cross-tenant superuser
visits and JWTs minted for foreign tenants no longer trigger silent
Owner grants. Two regression tests pin the new behaviour.
- service/user.synthFallbackMembership: default the synthesised role
to Viewer instead of Owner. The login response only feeds UI; the
backend re-derives the real role from tenant_members on every
request, so showing a Viewer UI is preferable to a misleading Owner
UI that would surface admin controls the backend will then 403.
- migrations/versioned/000043_tenant_rbac.down.sql: gate the destructive
drop behind a `weknora.allow_destructive_migration` GUC so an
accidental `migrate down` cannot wipe role data in production.
- frontend/src/stores/auth.ts: comment currentTenantRole as
rendering-only so future PRs cannot mistake the localStorage-backed
value for an authorisation signal.
API contract
- Memberships in LoginResponse and OIDCCallbackResponse now serialises
without omitempty, and buildMembershipsForUser / synthFallbackMembership
always return a non-nil slice — frontend can rely on JSON `[]`.
- handler/auth.GetCurrentUser includes memberships so /auth/me refreshes
recover currentTenantRole after a page reload.
Performance
- Add TenantService.GetTenantsByIDs (with the matching repo method) and
use it inside buildMembershipsForUser to batch tenant-name lookups
instead of issuing one query per membership.
Observability
- Replace the remaining `log.Printf` calls in middleware/auth with
contextual logger.Warnf / logger.Infof so role resolution events ride
the standard request-id / Langfuse trace plumbing. The successful
auto-promote line is tagged `[audit]` for log mining.
Tests
- internal/application/service/tenant_member_test.go: hand-rolled
fakeTenantMemberRepo plus 11 cases covering AddMember validation,
EnsureOwner idempotence, "cannot demote/remove the last Owner",
same-role no-op, ErrInvalidTenantRole / ErrMembershipNotFound, and
the TenantRole.HasPermission lattice.
- internal/middleware/auth_role_test.go: fakeMemberService plus 8 cases
pinning every branch of resolveTenantRole, including the two H1
regressions (cross-tenant superuser must not auto-promote; orphan
auto-promote requires the home tenant).
Lay the foundation for tenant-level RBAC without enforcing it yet.
Schema, types, middleware lookup, and login-flow shape changes ship now;
RequireRole/route annotations and member-management endpoints follow in
PR 2 and PR 3.
Highlights:
- New tenant_members table (Postgres + SQLite) with (user_id, tenant_id)
unique key and (owner|admin|contributor|viewer) role values. Migration
000043 backfills one row per existing user, marking the earliest-created
active user in each tenant as Owner and the rest as Contributor.
- KnowledgeBase.creator_id and CustomAgent.runnable_by_viewer added so
PR 2 can wire ownership-based and Viewer-runtime checks.
- Auth middleware loads the active TenantMember row, sets a new
TenantRoleContextKey, and falls back via three escape hatches:
cross-tenant superusers (CanAccessAllTenants) get Admin in the target
tenant; orphan tenants (zero members) auto-promote the first authenticating
human to Owner; and when tenant.enable_rbac is false (the v1 default)
unmatched lookups fail open with Admin to preserve current behaviour.
- Register / OIDC provisioning now insert an Owner membership for the
registrant via TenantMemberService.EnsureOwner.
- LoginResponse renamed tenant -> active_tenant and gained a memberships
array. POST /auth/switch-tenant issues a fresh token pair scoped to a
target membership; GET /auth/config exposes auth.registration_mode for
the frontend register-tab gating.
- New config keys: auth.registration_mode (self_serve | invite_only,
defaults to self_serve) and tenant.enable_rbac (defaults to false).
Both have WEKNORA_* env overrides.
- Frontend store persists memberships and exposes currentTenantRole;
Login.vue calls /auth/config and hides the Register tab in invite_only
mode. Go SDK types updated to match the new LoginResponse shape.
Wires KnowledgeBase.VectorStoreID and the ownership-aware retrieve factory
into the user-facing knowledge-base lifecycle:
- POST /knowledge-bases validates the requested vector_store_id against
the caller's tenant scope and the engine registry. New error codes
ErrVectorStoreBindingInvalid (2200) and ErrVectorStoreUnavailable (2201)
distinguish the typed branches without echoing UUIDs to the client.
- GET / POST / PUT / PUT-pin responses embed the bound store's display
metadata (name, source, engine_type, status) without exposing any
connection credentials. Cross-tenant shared KBs receive a suppressed
payload (vector_store_id stripped, source="shared") so operator-chosen
store names cannot be enumerated across tenants.
- POST /knowledge-bases/copy synchronously rejects clones whose target
has a different embedding model or vector store, before the async
clone task is enqueued. The async clone worker re-applies the same
checks for defense in depth.
- DELETE /vector-stores/:id refuses to remove a store with bound KBs,
inside a transaction that row-locks the store on PostgreSQL and
serializes via WAL on SQLite. unregister-from-registry is wrapped in
defer/recover so a panic surfaces as a structured warning instead of
silently leaking a stale engine.
- vector_store_id is immutable after creation. The GORM <-:create tag
blocks every ORM update path; the service-layer DTO omits the field
entirely; a reflection-based regression test catches any future
maintainer who adds it back to either layer.
- Empty-string vector_store_id is normalized to nil at both the create
path and inside SharesStoreWith, so rows persisted by callers that
did not run Normalize first cannot trip false same-store comparisons.
Part of #993. Depends on #994 and #1310.
Two encryption gaps and one fail-loud regression, fixed together because
they are inseparable in practice.
At-rest encryption gaps
-----------------------
Before this change, only Tenant.APIKey, Tenant.Credentials.WeKnoraCloud.AppSecret,
Model.Parameters.APIKey/AppSecret, and WebSearchProviderParameters.APIKey
were encrypted with SYSTEM_AES_KEY. MCP and DataSource were missed:
- MCPAuthConfig.Value/Scan went through bare json.Marshal/Unmarshal, so
api_key and token were written to mcp_services.auth_config (json) in
plaintext regardless of whether SYSTEM_AES_KEY was set.
- DataSource.Config (types.JSON) is a byte passthrough, so connector
secrets (Notion token, GitHub PAT, Feishu app_secret, ...) inside the
Credentials map were stored in plaintext too.
Now both follow the ModelParameters / WebSearchProviderParameters pattern.
Encryption operates on a local copy to avoid mutating the caller's
in-memory struct (otherwise the next read of the same *MCPAuthConfig
would surface ciphertext as if it were the API key). For DataSource the
ToJSON ↔ ParseConfig pair is the only application boundary that
materializes credentials, so encrypting there is sufficient and avoids
having to fork the generic types.JSON.
Backward compatibility
----------------------
DecryptStoredSecret returns rows without the enc:v1: prefix verbatim, so
historical plaintext records keep working. Next write re-saves them
encrypted. No migration step.
Graceful Scan-time degradation
------------------------------
The strict "fail loud on key rotation" behaviour of DecryptStoredSecret
turned into a footgun once it ran from GORM Scan: a single broken row
caused ListModels / ListMCPServices / ListDataSources to return zero
results, hiding all other rows and giving the operator no way to find
which credential needed re-entering.
New utils.DecryptStoredSecretLenient: same logic, but on failure returns
("", false). The strict variant is kept for hot paths that are about to
actually USE the credential (so a rotated key surfaces loudly there
rather than silently degrading to "missing api key").
All Scan paths switch to lenient + per-failure log line:
ModelParameters.Scan, WebSearchProviderParameters.Scan,
MCPAuthConfig.Scan, DataSource.ParseConfig, CredentialsConfig.Scan.
On lenient failure the field blanks out (never ciphertext), the row
still loads, the DTO reports configured=false so the UI shows "not
configured", and any upstream call attempted with the empty value fails
with a clear "missing api key" rather than sending ciphertext as the key.
Tenant.AfterFind (Tenant.APIKey) intentionally stays strict — it is the
primary HTTP auth credential and a silent empty value would change auth
semantics ambiguously.
Replace the legacy "redacted placeholder + Clear* boolean" pattern with
dedicated per-resource credential subresources across MCP services,
Models, Web search providers, and Data sources.
Why
---
The previous design had three problems:
1. Main PUT body carried secret fields. The frontend echoed back a
redacted "***" placeholder, and a fragile MergeUpdate / IsRedactedOrEmpty
defense in the service layer tried to detect "user did not change this"
vs "user wants to clear this". A regression in that defense (or a new
frontend forgetting it) silently overwrites the stored secret with the
placeholder.
2. The "remove this credential" UX was a red checkbox under a pre-filled
password input. Three intents (preserve / replace / clear) collapsed
onto one field, and credential changes were bundled with unrelated
config edits in the same submit. Users wiped working keys by mistake.
3. Secret presence was inferred from "did the response come back with a
'***' placeholder", which couples the contract to a magic string.
Design
------
Each of the four resources now exposes:
PUT /{resource}/{id}/credentials # write one or more fields
DELETE /{resource}/{id}/credentials/{field} # clear a single field
"Is this configured?" metadata travels on the main resource response as
a typed map (dto.MCPServiceResponse.Credentials etc.) — no separate GET
endpoint. The frontend reads the configured boolean from the main GET
and never sees secret values at all.
Main PUT endpoints now ignore any secret fields in the request body and
log a deprecation warning if they appear, so legacy clients fail loudly
rather than overwriting silently.
Frontend
--------
- New reusable <CredentialResource> component renders a three-state card
(unconfigured / configured / editing) and drives the dedicated
endpoints. Used by MCP, Model, Web search; DataSource has a bespoke
card with the same behaviour because its credentials are a single
atomic per-connector map.
- Cancel from edit mode now restores state synchronously from the meta
prop. The previous async refresh() was a no-op while state was still
'editing', leaving the input frozen open.
- Remove is single-click + toast. The danger-themed button is the
deterrent; a modal confirm adds friction without adding safety (the
plaintext is irrecoverable client-side either way — recovery means
re-typing).
DTOs (internal/handler/dto/) are deliberately separate from the GORM
models so "no secret in response" is a compile-time invariant: a future
contributor cannot leak a secret without explicitly adding the field to
the DTO, which is review-able in a single diff.
Storage is unchanged — credentials still live in the existing jsonb /
parameters columns. No schema migration.
Cleanup
-------
- types.MCPService / types.Model / types.WebSearchProviderEntity /
types.DataSourceConfig: drop ClearAPIKey / ClearToken / ClearAppSecret
/ ClearCredentials boolean fields, MergeUpdate(), RedactSensitiveData().
- utils/types/secret.go: drop PreserveIfRedacted / IsRedactedOrEmpty.
RedactedSecretPlaceholder constant is retained because VectorStore
still uses the old pattern and is out of scope here.
- Frontend hasExistingApiKey / clearApiKey / convertToLegacyFormat
redaction handling removed; i18n keys renamed secret -> credential.
Closes#988
Unify the "write-only secrets" pattern across MCP services, Models,
WebSearch providers, and DataSources. Fix two related defects:
1. GET responses leaked plaintext credentials — GetMCPServiceByID
while List masked, all Model / WebSearchProvider / DataSource GET
responses unconditionally.
2. When the frontend echoed a masked placeholder back on save, the
update path overwrote the stored secret with the mask — silently
wiping the real credential. MCP surfaced this most visibly:
ListMCPServices masked Token to "first4****last4" and
McpServiceDialog round-tripped that masked string back into the DB.
- internal/types/secret.go introduces the shared sentinel
RedactedSecretPlaceholder = "***" and two helpers:
- IsRedactedOrEmpty(s): true for "" or the sentinel
- PreserveIfRedacted(incoming, existing): merge predicate
- Each resource adds a RedactSensitiveData (mutates-in-place) or
Redacted (returns-a-copy) method and a MergeUpdate contract:
empty or placeholder preserves, any other value replaces, and a
write-only Clear* bool flag (gorm:"-", json omitempty) wipes with
precedence over preserve / replace. Clear* never persists.
- Every GET / List / Create / Update response path now redacts.
VectorStore's pre-existing "***" literal is refactored onto the
shared constant for consistency.
All four dialogs — McpServiceDialog, ModelEditorDialog,
WebSearchSettings, DataSourceEditorDialog — never pre-fill secret
inputs. The stored-state signal is collapsed into the input
placeholder: "•••••••• (Enter new value to replace)" when the server
reports a stored value, falling back to the connector-specific
placeholder otherwise. The input itself carries the "something is
there" cue, matching how password dots render.
An opt-in "Remove this credential" checkbox appears in edit mode when
a value is stored; checking it disables the input and triggers a
danger-themed confirmation dialog before submit. DataSource rolls the
clear up to a connector-level "Remove all credentials" flag since its
credentials are per-connector atomic bundles (e.g. Feishu app_id +
app_secret).
i18n updates across en-US, zh-CN, ko-KR, ru-RU:
secret.storedPlaceholder / clearHint / confirmClearTitle /
confirmClearBody.
- MCP update handler's manual map-to-struct projection was silently
dropping clear_api_key / clear_token / custom_headers on partial
updates. Pull them off the map explicitly.
- MCP configChanged detection now snapshots URL / StdioConfig /
TransportType before the merge so connection recycling still
triggers on real transport changes (not just whenever AuthConfig
is present).
- MCP CustomHeaders: nil preserves existing, non-nil replaces. Was
always overwriting, silently wiping headers on any auth-bearing
update.
- GetMCPServiceByID now redacts (was plaintext while List masked).
- WebSearchProvider update preserves Name / Description when empty
instead of clobbering them.
- WebSearchProvider TestProviderRaw rejects the "***" placeholder up
front with an actionable message rather than forwarding it to the
upstream provider API as a literal key.
- Model builtin hideSensitiveInfo now preserves non-secret metadata
(SupportsVision, Provider, ExtraConfig, InterfaceType,
ParameterSize, EmbeddingParameters) that were previously zeroed
out for builtin model cards.
- Replace local equalStringSlices / equalStringMaps helpers with
stdlib slices.Equal / maps.Equal (Go 1.21+).
- internal/types/{secret,mcp,model,web_search_provider,datasource}_test.go:
table-driven coverage of merge, redact, clear, and write-only flag
invariants for every resource.
- internal/application/service/mcp_service_test.go: full service layer
with a fake repo; covers ListMCPServices / GetMCPServiceByID
redaction, UpdateMCPService partial-update honoring of Clear*,
CustomHeaders preserve-on-nil / replace-on-non-nil, and
clear-flag precedence over submitted values.
Additive only. No changes to request or response JSON shape beyond
the write-only Clear* flags (omitted by default via omitempty). No DB
schema changes. No SDK struct field removals. Existing clients that
don't know about the new flags continue to function unchanged.
- DataSource.Config jsonb at-rest encryption. Pre-existing residual
risk; tracked as a follow-up.
- Renaming VectorStore.MaskSensitiveFields to match the new
RedactSensitiveData / Redacted naming convention. Left alone to
minimize diff; a follow-up rename PR can unify.
Extracts the 25 repeated NewCompositeRetrieveEngine call sites across
seven services into two factory functions with tenant-ownership
verification: CreateRetrieveEngineForKB (synchronous) and
CreateRetrieveEngineFromPayload (async task handlers). Promotes
GetByStoreID from concrete-only to interfaces.RetrieveEngineRegistry.
Extends KBDeletePayload / IndexDeletePayload with an omitempty
VectorStoreID snapshot.
Design highlights
- Factory verifies tenant ownership of the resolved store
(defense-in-depth) so a gap in PR3 validation or a tampered Asynq
payload cannot cross tenants. Cross-tenant attempts return
ErrVectorStoreForbidden with a structured log entry.
- Sentinel errors (ErrTenantInfoMissing, ErrVectorStoreNotFound,
ErrVectorStoreForbidden) let async handlers classify non-retryable
failures as asynq.SkipRetry.
- nil and empty-string pointers for vectorStoreID normalize to
"unbound" so callers never send an empty UUID to GetByStoreID.
- resolveBoundEngine constructs engineInfos directly (with
slices.Clone of Support()) instead of going through
NewCompositeRetrieveEngine. The latter only reads from the
byEngineType map and cannot reach DB stores in byStoreID -- multiple
stores can share the same engine type. Unbound fallbacks still reuse
NewCompositeRetrieveEngine so the tenant effective-engines path is
byte-for-byte unchanged.
- DeleteKnowledgeBase now reads the KB before soft-delete so the
enqueued KBDeletePayload carries a VectorStoreID snapshot; once the
row is soft-deleted GORM's default scope hides it. Cost: +1 SELECT
per KB-delete request (a rare admin operation).
- enqueueIndexDeleteTask gains a vectorStoreID *string parameter that
the single caller (tag.go) populates from the owning KB. The worker
(ProcessIndexDelete) validates ownership via the factory.
- cleanupKnowledgeResources loads the KB so bound-store cleanup routes
correctly. If the load fails it warns and falls back to tenant
effective engines; orphan vectors become observable in logs instead
of disappearing silently.
- DeleteKnowledges (batch) stays on the tenant effective-engines path
-- batches can span stores and multi-store fan-out is PR4 scope.
Noted inline.
Merge safety
All knowledge bases have VectorStoreID = NULL at merge time (introduced
nullable in PR1, Tencent/WeKnora#994). Every factory call falls back to
NewCompositeRetrieveEngine(registry, tenantInfo.GetEffectiveEngines()),
which is the existing code path for these KBs. Async payloads from
before this change decode to VectorStoreID = nil (omitempty) and take
the same fallback.
Testing
- go build ./...
- go vet ./...
- go test -race -count=1 ./internal/application/service/retriever/...
covers unbound (nil/empty), bound, cross-tenant, unregistered store,
ownership infra error, legacy payload shapes (missing field/null/
empty string/nil), tampered payload, and parallel invocation.
- grep -r NewCompositeRetrieveEngine internal/application/service
returns zero results outside factory.go/composite.go.
Refs: Part of Tencent/WeKnora#993. Depends on Tencent/WeKnora#994.
MiMo and DeepSeek V3.2/V4 reject multi-turn requests in thinking mode
when the prior assistant message lacks reasoning_content with HTTP 400:
"The reasoning_content in the thinking mode must be passed back to the API."
Agent's ReAct loop is the worst-case scenario — every round produces tool
calls, exactly the case DeepSeek's docs specify reasoning_content MUST
participate in subsequent context.
Plumb reasoning_content through the full assistant-message round-trip:
- chat.Message / types.ChatResponse / types.AgentStep gain a
reasoning_content field (AgentStep persists via the existing
Message.AgentSteps jsonb column, no migration needed).
- streamLLMToEventBus accumulates reasoning chunks into
result.ReasoningContent and surfaces it on the round's ChatResponse.
- engine.runReActIteration writes it onto AgentStep so cross-turn replay
preserves it.
- observe.appendToolResults attaches it to the same-turn assistant
message; agent_history.buildAssistantHistoryMessages does the same on
cross-turn replay.
- RemoteAPIChat.ConvertMessages forwards it on assistant turns to
openai.ChatCompletionMessage.ReasoningContent (already supported by
go-openai); providers that don't recognize the field ignore it.
Tests cover the three boundaries: ConvertMessages serializes it for
upstream, appendToolResults preserves it within the same turn, and
buildAssistantHistoryMessages replays it across turns.
Scope is intentionally limited to Agent mode — KnowledgeQA's chat
pipeline and Anthropic's signed thinking_blocks are separate fixes that
require schema changes (rendered_content / thinking_blocks columns).
Fixes#1302
Quick-answer (RAG) agents can now configure a separate chat model for
the query-understanding stage (rewrite + intent classification),
decoupling it from the main conversation model so users can route the
lightweight rewrite call to a cheaper / faster model.
- Add CustomAgentConfig.QueryUnderstandModelID and plumb it through
PipelineRequest and ChatManage.Clone.
- query_understand plugin prefers QueryUnderstandModelID on the
text-only path; falls back to ChatModelID if the configured model
cannot be resolved (with a warn log). Multimodal path is unchanged
to keep vision-capable model selection intact.
- AgentEditorModal exposes a ModelSelector under the existing Query
Rewrite block; empty means reuse the main chat model.
- Add i18n strings (queryUnderstandModel / placeholder / desc) for
zh-CN, en-US, ko-KR, ru-RU.
- Extend Go SDK AgentConfig with the new field.
The legacy in-pipeline DuckDB SQL data-analysis stage used to run on every
quick-answer RAG request whose retrieved chunks included a CSV/Excel file,
adding one extra LLM round-trip (~3s) to generate a SQL query that most
plain Q&A users never wanted. There was no way to disable it.
Introduce a per-agent DataAnalysisEnabled flag (default off), wire it
through PipelineRequest, and gate the DATA_ANALYSIS stage on it. Surface
the toggle in the agent editor for quick-answer agents with at least one
knowledge base attached.
Refs: https://github.com/Tencent/WeKnora/issues/1244
Add three new optional filters to the document list under a knowledge base
detail page — parse status, source/channel, and updated time range — and
rework multi-select to no longer cause the card title to jitter on hover.
Backend
- Introduce types.KnowledgeListFilter to aggregate optional filter dimensions
(tag, keyword, file_type, parse_status, source, updated_from/to) and switch
ListPagedKnowledgeByKnowledgeBaseID (repository/service/interface) to accept
it instead of a growing positional parameter list.
- The ListKnowledge HTTP handler accepts new parse_status, source, start_time
and end_time query params; time params accept RFC3339, "YYYY-MM-DD HH:MM:SS"
and "YYYY-MM-DD". The repository routes source="manual"/"url" onto the type
column to stay consistent with file_type semantics; other source values match
the channel column.
- Update the four other callers (agent_service, initialization) to pass an
empty filter struct, preserving prior behavior.
Frontend
- Add three controls in the doc-filter-bar (status select, source select,
date-range picker with future-date disabled) wired through getKnowled /
listKnowledgeFiles into the new backend params.
- Replace the hover-triggered card checkbox with an explicit "批量管理" mode
(mirrors the session list UX): in card view the checkbox only renders while
batch mode is on, entered via the per-card "..." menu; the list view keeps
its leading checkbox column. Switching from list to grid auto-enables batch
mode when something is already selected, so the selection stays visible.
- DocumentBatchBar now stays open whenever batch mode or selection > 0, and
its "取消选择" button both clears the selection and exits batch mode.
API surface sync
- Regenerate Swagger artifacts (docs/docs.go / swagger.json / swagger.yaml).
- Update docs/api/knowledge.md with the new query parameters.
- Add backward-compatible ListKnowledgeWithFilter + KnowledgeListFilter to the
Go SDK; the existing ListKnowledge keeps its signature.
i18n
- New filter labels in zh-CN / en-US / ko-KR / ru-RU; reuse existing
menu.batchManage / batchManage.cancel for the multi-select strings.
The Agent ran a parallel context cache (Redis/in-memory) on top of the
messages table to feed multi-turn history into the LLM. That dual-write
caused subtle drift (e.g. compression diverging from DB, system-prompt
swaps lost on restart) and required a separate ClearContext path on
session/IM clear.
Make the messages table the single source of truth:
- Add service.LoadAgentHistory: rebuilds chat.Message history per turn
from the persisted messages, expanding AgentSteps into proper OpenAI
assistant_with_tool_calls + tool messages and replaying the canonical
final answer (with <think> blocks stripped). final_answer tool calls
are filtered to avoid duplicating the trailing answer.
- Make AgentEngine stateless across turns: drop ContextManager / sessionID
cache plumbing from the engine, agent service, and CreateAgentEngine
signature. The engine only uses sessionID for logging/event emission.
- Wire AgentQA to load history from DB on demand using HistoryTurns
(default 5) when MultiTurnEnabled, otherwise run with empty history.
- Delete the llmcontext package (ContextManager interface, Redis/memory
storage, factory) and the SessionService.ClearContext API path; IM
/clear and session message clear no longer need to invalidate cache.
Behavior preserved: KnowledgeQA-mode replay is unchanged (turns with
empty AgentSteps just produce the single canonical assistant message),
and Agent-mode turns now consistently see prior tool calls and results.
Add an opt-in human approval gate so Agent runs pause before executing
MCP tools that operators flag as dangerous, surface an approval card in
the chat UI, and only resume after the user approves (optionally with
edited args) or rejects.
Backend
- New mcp_tool_approvals table + repo/service to mark per-tool approval
required (PG migration 000042 + sqlite init).
- approval.Gate coordinates RequestAndWait / Resolve with sync.Once
delivery, configurable timeout, and Redis Pub/Sub fan-out so multi-
replica deployments work without sticky sessions.
- MCPTool.Execute integrates the gate; uses a round-level ApprovalCtx
(without the per-tool 60s timeout) for the wait, and re-derives a
fresh 60s exec ctx after approval so CallTool keeps a full window.
- New SSE response types (tool_approval_required / _resolved) and
EventBus events plumb approval state to AgentStreamDisplay.
- REST: list/set per-tool approval flag, resolve pending approval.
- Configurable via agent.tool_approval_timeout_seconds (yaml) or
WEKNORA_AGENT_TOOL_APPROVAL_TIMEOUT env (accepts seconds or Go
duration).
Frontend
- MCP settings: per-tool "require approval" switch on the test panel.
- Chat: ToolApprovalCard renders the pause point with editable JSON
args, validation feedback, mm:ss countdown that turns warning/danger
near deadline, and a resolved state that retains context.
- i18n strings added for zh-CN / en-US / ko-KR / ru-RU.
Docs
- docs/zh/mcp-approval.md covering behavior, config, API, deployment
considerations (Redis cross-instance, restart limitations).
The wiki ingest post-process pipeline OOM'd and ran for hours on KBs
with ~40k documents. The dominant tail was a per-batch ListAllPages
that pulled every page (multi-MB content blobs) into Go memory, plus a
24h-TTL Redis pending list whose data could be evicted before a long
serialized backlog drained. dedup did O(P × N) Jaccard scoring in Go
on every batch. Lint loaded the full graph. The index page kept the
entire wiki directory in its content column and rewrote the TOAST
chunk on every ingest. None of these survive at 4w docs.
This change reworks the write path end to end and pulls the durable-
queue + dead-letter primitives out of wiki and into shared
infrastructure that every asynq task type now benefits from.
- migration 000041: task_pending_ops + task_dead_letters tables, plus
three GIN indexes on wiki_pages (source_refs jsonb_path_ops,
source_refs text fallback, lower(title) trgm).
- TaskPendingOpsRepository + TaskDeadLetterRepository in
internal/application/repository/task_queue.go: cursor-paginated
list, atomic IncrFailCount via UPDATE…RETURNING, dedup-key scoped
delete that refuses empty keys so a buggy caller can't wipe a KB's
queue.
- internal/middleware/asynqdl: writes a task_dead_letters row when an
asynq task exhausts its retry budget. Payload-agnostic — a small
probe struct extracts TenantID + scope hints across every existing
payload type, so summary:generation / image:multimodal /
faq_import / kb:clone / etc. all dead-letter without per-handler
code. Best-effort: an insert failure never masks the underlying
task error. Installed in router/task.go before the langfuse mw so
it sees raw errors.
- Pending queue moved Redis → PG (no TTL, restart-durable). Redis
keeps just the active-batch lock and the delete tombstone.
- In-batch retry budget (wikiMaxFailRetries=5) tracked via
pendingRepo.IncrFailCount; over the cap the row is moved to
task_dead_letters with task_type=wiki:ingest, related_id=knowledge
id. Asynq retries (10) are handled by the new middleware.
- Removed the per-batch ListAllPages. WikiBatchContext now carries
lazy fetcher closures (SlugTitleMany, SummaryByKnowledgeID) with
mutex-protected caches; reduce reaches for titles / summaries on
demand instead of pre-loading the whole KB.
- getExistingPageSlugsForKnowledge now uses ListSlugsBySourceRef,
which the new GIN index on source_refs serves as a Bitmap Index
Scan instead of a sequential text LIKE.
- Dedup pre-filter uses pg_trgm via FindSimilarPages
(idx_wiki_pages_title_trgm). For each new entity/concept (and each
of its aliases) we ask the DB for the top-K trigram-similar
existing pages and union the results — bounded prompt size, no Go-
side O(P × M) loop. Small KBs (≤25 entities) bypass the pre-filter.
- cleanDeadLinks / injectCrossLinks are scoped to the batch's
affected slugs (a few dozen) instead of every page in the KB. Both
use the new lite ListBySlugs / ExistsSlugs repo methods so they
pull only slug + title + outlinks, not full content.
- Slug fuzzy resolve (slug_fuzzy.go): when the LLM emits
[[bad-slug|display]] the cleanup path now tries display-text
reverse lookup → hyphen/case normalized equality → char-bigram
Jaccard ≥ 0.8 before stripping. Recovers the common pinyin-word-
break drift case ("shang-hai-tower" vs "shanghai-tower") in place
instead of replacing the link with plain text.
- rebuildIndexPage uses ListByTypeRecent(200) for the first-time
intro and drops the full DocumentSummaries blob from the
incremental update prompt, so its context stays bounded regardless
of KB size.
- Concurrency tunables surfaced in WikiConfig: IngestBatchSize /
IngestMapParallel / IngestReduceParallel, with sensible defaults
via OrDefault helpers. scheduleFollowUp drops to ProcessIn(0) so
follow-ups don't waste asynq retry slots bouncing on the active
lock.
- RunLint walks pages via the new ListPagesCursor in 200-page
windows and computes the live-slug set with a one-column
ListAllSlugs Pluck instead of a Limit:0 GetGraph that materialized
every node + edge. Memory is now bounded; 40k pages walks in
constant ~4MB.
- 14 GORM tests for both repos against an in-memory SQLite mirror of
the production DDL (task_queue_test.go).
- 6 tests for the dead-letter middleware covering retry budget
detection, payload-agnostic scope inference, error truncation, and
repo-failure isolation.
- 7 tests for the slug fuzzy resolve helper covering the three
resolution stages, display-text priority over normalized equality,
bigram fallback acceptance, and rejection of unrelated slugs.
- Existing wiki_ingest / wiki_lint / wiki_page tests updated to the
new fetcher / cursor APIs.
Fixes O(n²) write amplification during wiki ingest on large KBs. Previously
every ingest/retract op re-wrote the single `slug='log'` row end-to-end and
every batch re-wrote the entire `slug='index'` directory markdown. On a 40k-
doc KB the log row grew to tens of MB and the index row to several MB, so
each batch triggered giant TOAST updates that dominated ingest wall time.
Log: new `wiki_log_entries` event table (`id DESC` indexed per KB) replaces
the single TEXT row. Batch ingest now collects entries and flushes them
once per batch via `AppendBatch`. Each entry stores `pages_affected` as
JSONB `[{slug,title}]` so the UI can render real titles; custom Scan falls
back to legacy `[]string` so older rows still deserialize.
Index: `wiki_pages[slug=index].content` keeps only the LLM-generated intro
(a few KB). The directory is now served by a structured paginated API
(`GetIndexView`) that reads `slug/title/summary` per type with cursor
pagination, so the agent and the frontend only pull the slice they need.
`RebuildIndexPage` degrades to a no-op; agent `wiki_read_page('index')`
synthesizes a small top-K overview and points callers at `wiki_search`.
Ingest resilience: LLM calls wrap with 3-attempt exponential backoff on
transient errors (5xx/408/429, transport resets/timeouts). Summary/extract
failures now bubble up so the batch's failed-op requeue path runs instead
of silently dropping the doc.
Frontend: sidebar Index/Log entries switch to dedicated views. Index view
streams intro → Summary → Entity → Concept → Synthesis → Comparison via
IntersectionObserver (with a nextTick re-check so small KBs still load
every section). Log view uses cursor pagination; pages_affected renders
titles with slug tooltip.
- new migration: migrations/versioned/000040_wiki_log_entries.{up,down}.sql
- tests: log repo pagination + legacy Scan, ListByTypeLight windowing,
renderIndexOverviewForAgent output, isTransientLLMError classifier
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wikis with tens of thousands of pages used to crash the browser trying
to render the entire link graph at once (30MB+ JSON, 100k+ SVG elements).
This change moves the graph viewer from "fetch everything, render
everything" to "fetch a slice, expand on demand" with several matching
UX improvements.
Backend
* GET /wiki/graph now accepts `mode` (overview | ego), `center`, `depth`,
`types`, `limit` query params. Default is overview top-500 by
link_count, capped at 2000. Response includes a Meta object so the
frontend can render a truncation hint and drive the UI.
* Pure helper `computeGraphSubset` extracted for testability; six unit
tests cover overview truncation / type filter / ego BFS / missing
center error.
* WikiLintService passes Limit=0 (uncapped) so link integrity checks
still walk every page.
* Repo `Search` adds CASE-based relevance ranking (title 4, slug 3,
summary 2, content 1) so full-text results put the most obvious
matches first instead of whatever was updated most recently.
Frontend
* Graph viewer fetches overview on entry; ego pivot on double-click,
search, wiki-link click, URL ?slug=, global issues jump.
* Shift+click or hover ⊕ button blooms neighbors onto the current
canvas additively. Bloom tracks generations and LRU-evicts oldest
when total exceeds 1500 nodes; ego center / selected node / latest
anchor are always protected.
* "Grow frontier (N)" legend action expands every dashed-ring node in
one click (6-way concurrency) while skipping Index/Log super-nodes
that would otherwise dump the whole wiki onto the canvas.
* Node dashed expansion ring + drawer "X/Y neighbors shown" hint tell
users which nodes still have neighbors to load.
* Search dropdown uses remote full-text search (debounced, sequence-
numbered to drop stale responses) and falls back to the overview
top-500 snapshot when the input is empty.
* Type filters now round-trip to the server so top-N is always
computed from the user's active type set rather than hiding nodes
client-side and shrinking the view.
* Status card replaces the cramped "centered on X · N hops · M nodes"
line with a structured focus/overview summary, resolving slugs to
page titles.
* Help popover in the legend lists every canvas shortcut.
i18n entries added for zh-CN / en-US / ko-KR / ru-RU.
Heuristic and recursive splitters preserve original byte positions in
Chunk.Content (the End-Start == RuneCountInString invariant is required
by document-reconstruction code paths), which means a chunk sliced at a
boundary often carries leading/trailing newlines from the boundary
itself. Feeding that whitespace into the embedding model dilutes the
vector and wastes tokens for no benefit.
TrimSpace the body inside EmbeddingContent for both chunker.Chunk and
the two types.Chunk / types.ParsedChunk mirrors. Inner whitespace is
preserved; positions/Content are unchanged.
Also clean up appendChunk: the previous code computed a trimmed string
for the empty-check then stored a separate untrimmed copy, which read
as if it might be intentionally divergent. Replace with an explicit
'raw text, skip if pure whitespace' shape and document the invariant.