Commit Graph

2611 Commits (0827cd6cf02ba6daedc8c9cd9762bf93abf2ee7a)

Author SHA1 Message Date
ttommybot 0827cd6cf0 fix: retry document summaries before fallback 2026-08-05 19:29:32 +08:00
01luyicheng 17372215a2 docs(mcp-server): document create_knowledge_from_file/text tools
List create_knowledge_from_file (previously missing from the README tool
list) and the new create_knowledge_from_text in README.md and EXAMPLES.md,
and add a CHANGELOG entry under Unreleased.
2026-08-05 19:25:25 +08:00
01luyicheng 54c4309c65 feat(mcp-server): add create_knowledge_from_text tool
Add an MCP tool that creates a knowledge entry from raw Markdown text by
calling the existing backend endpoint
POST /knowledge-bases/{id}/knowledge/manual (handler.CreateManualKnowledge).
This completes the "text" half of #323, whose "file" half was already
covered by create_knowledge_from_file.

The tool defaults to status="publish" so the created entry is chunked,
embedded and made searchable immediately — matching the API/MCP use case
where there is no UI to publish a draft. Callers may pass status="draft"
to save without indexing.

- New client method WeKnoraClient.create_knowledge_from_text.
- New @mcp.tool() create_knowledge_from_text, resolving kb_id by name or
  UUID (consistent with create_session / hybrid_search).
- Add focused unittests verifying endpoint, request body shape, status
  default/override, and kb_id resolution wiring.
- Bump the stdio tools-list count assertion in test_mcp_transports.py
  from 29 to 30 for the newly registered tool.
2026-08-05 19:25:25 +08:00
xiaowang 520be78766 fix(chunking): normalize line endings before splitting / 在分块前统一换行符 2026-08-05 19:24:45 +08:00
wizardchen 63b5f353c1 Improve Wiki reader layout and revision history UX.
Restructure the page header and footer for clearer metadata, add field-level revision diffs with correct incremental comparison (including v1 initial content), and align footer links with in-body wiki link styling.
2026-08-05 18:15:11 +08:00
wizardchen 03d6f90c5a fix(faq): prevent duplicate entries when a create is retried
A caller that resends CreateFAQEntry before the first call answers could
create several rows for the same standard question. Indexing runs inline and
the embedding call retries with exponential backoff, so one create can take
far longer than an impatient caller waits; meanwhile the duplicate-question
check only looked at ChunkStatusIndexed, so a sibling create still mid-flight
was invisible and every retry passed the check.

- Serialize concurrent creates of the same question behind a Redis SetNX
  guard, keyed by tenant + knowledge base + question hash, with an in-process
  fallback when Redis is unavailable so an outage cannot block FAQ authoring.
- Widen the duplicate check to every non-deleted status. This also fixes
  ChunkStatusDefault entries, which are listed to users but never blocked
  duplicates.
- Bound the indexing step of an interactive create so a degraded embedding
  service cannot stretch the request out; bulk and background indexing keep
  the full retry budget.
- Report a failed rollback instead of discarding the error, since a leftover
  stored chunk is hidden from the list yet now matches the duplicate check.
2026-08-05 17:05:51 +08:00
wizardchen 1985a59b69 fix(chunking): sync preview swagger and normalize splitter config
Centralize ingestion base splitter defaults in chunker.NormalizeSplitterConfig
so the preview endpoint applies the same normalization as upload. Regenerate
Swagger for parent-child preview fields and add regression tests for single-
level preview, default parent/child sizes, and context headers.
2026-08-05 16:17:48 +08:00
wizardchen 39f4220c13 fix(wiki): ensure prompt language is always resolved
Wiki generation sometimes rendered "Write in ." because async workers
read empty locale payloads without falling back to context or defaults.
Unify language resolution helpers and apply them at enqueue and consume
time so every wiki and related async prompt gets a non-empty language.
2026-08-05 16:14:50 +08:00
xiaowang d2888e9565 fix(chunking): align preview with parent-child splitting / 修复预览与父子分块不一致 2026-08-05 15:09:14 +08:00
wizardchen b191201f44 style(client): gofmt message.go 2026-08-05 14:42:51 +08:00
wizardchen e71c2dce2e fix(api): address resource_urls review follow-ups
Harden public URL mode: stop logging signed URLs at INFO, copy messages
before rewriting history responses, reuse winning grants on insert races,
sync the Go client resource_urls query param, and add knowledge-search
coverage.
2026-08-05 14:42:51 +08:00
wizardchen eb6ac0110c fix(api): close the resource_urls=public gaps found in review
Follow-up on resource_urls=public. Review turned up one authorization gap plus a
set of correctness problems in the streaming path.

Authorization:

  - Embed channels were documented as excluded, but nothing enforced it. The
    delegating handlers hand the visitor's own gin context to the shared
    handlers, so ?resource_urls=public — and a deployment-wide
    RESOURCE_URL_MODE=public — reached anonymous traffic and produced
    shareable, credential-free URLs that bypass the channel-scoped
    /embed/:channel_id/files proxy. ensureEmbedSession now pins the request
    context to handle mode, which covers every delegated endpoint including
    future ones. The downgrade is silent so a client that forwards the
    parameter keeps working.

  - A knowledge-base-restricted API key is refused with 403. Such a key is
    already denied /files because a raw storage path cannot be bound to its
    allow-list, so handing it anonymous file URLs would reopen that hole from
    the other side. The default handle mode stays available to it.

Streaming:

  - Rewriter.ref holds its lock across resolution. Resolver implementations keep
    an unsynchronised per-provider cache, so the previous unlocked window was a
    real data race for any caller sharing a StreamRewriter between goroutines —
    which its own doc comment invited. Serialising also collapses a concurrent
    duplicate into one signature.

  - Release the holdback buffer on every path that ends a stream while the
    client is still connected: a user-requested stop, an error event, and
    giving up on the event store. Only completion flushed before, so the tail
    generated just before a stop was silently dropped.

  - A released tail carries the metadata of the event it was cut from instead of
    a bare event_id.

  - The maxHeldBytes safety valve aligns to a UTF-8 boundary rather than cutting
    a character in half, and an unclosed "![…](" only holds back a
    whitespace-free destination under 2048 bytes, so prose that merely mentions
    Markdown image syntax no longer stalls the stream behind the buffer.

Also resolves the mode before running a knowledge search instead of after, so an
invalid value no longer costs a full retrieval, and warns once per distinct bad
RESOURCE_URL_MODE value instead of once per process.

Co-authored-by: lyingbug <lyingbug@users.noreply.github.com>
2026-08-05 14:42:51 +08:00
wizardchen 8b00cb71a3 perf(storage): reuse a live resource access grant per resource
Resolving a resource:// handle into a public URL inserted one capability row
every time. Rendering an answer resolves the same image across many chunks and
re-reading a message history resolves it again on every call, so a read endpoint
behaved like a write-heavy one.

The table stores only sha256(token), so an existing grant cannot be reused by
looking it up — the plaintext token is unrecoverable. Derive the token instead:
HMAC-SHA256(SYSTEM_AES_KEY, resource id + time window). The same resource in the
same window derives the same token, so the row can be found by its hash and
reused; only the first request in a window inserts. The window is half the TTL,
so a reused grant always has at least TTL/2 of life left.

This keeps the storage properties intact: the table still holds only hashes and
the plaintext cannot be reconstructed from a dump without SYSTEM_AES_KEY, while
authorization stays entirely on the row, so expiry and revocation still decide
whether a token resolves. Without SYSTEM_AES_KEY, grants stay random and
per-request as before. A stable URL per window also lets client and CDN caches
hit.

Stop deleting revoked grants during cleanup. A revoked row has to survive until
its expiry, otherwise a token derived for the same resource and window would
re-create it and revive the access that was just revoked. grant.revoked_at has
no writer yet, so this changes nothing today, but the derived token makes the
tombstone load-bearing.

Co-authored-by: lyingbug <lyingbug@users.noreply.github.com>
2026-08-05 14:42:51 +08:00
lyingbug 061319581d refactor(chat): build SSE payload after the liveness check
In public resource URL mode, building the payload consumes the chunk into the
holdback buffer, so returning between build and write would drop it. Reusing
emitStreamEvent also removes the duplicated flush-before-completion logic.

Co-authored-by: lyingbug <lyingbug@users.noreply.github.com>
2026-08-05 14:42:51 +08:00
lyingbug ef1f0b36c4 docs(api): document resource_urls and RESOURCE_URL_MODE
Adds a cross-cutting "文件与图片引用" section to docs/api/README.md covering the
parameter, the deployment default, the endpoints it applies to, and the two
caveats that matter in practice: it needs APP_EXTERNAL_URL (or a publicly
reachable storage backend) to produce a link at all, and the links it produces
are time-limited but anonymously readable.

Also annotates the affected endpoints for Swagger and refreshes the generated
files for those parameters only, leaving unrelated drift in the committed
swagger output alone.

Co-authored-by: lyingbug <lyingbug@users.noreply.github.com>
2026-08-05 14:42:51 +08:00
lyingbug 96e28b2438 test(api): cover resource_urls=public across the rewriter, handlers and SSE wire
Includes an end-to-end ContinueStream test that asserts the full SSE wire
output. It caught a leak: the references event carries its results twice, and
the copy in Data holds the typed []*SearchResult that CopyData did not
traverse, so handles reached a caller that asked for public URLs. CopyData now
handles the typed slice, []string and map[string]string forms.

Co-authored-by: lyingbug <lyingbug@users.noreply.github.com>
2026-08-05 14:42:51 +08:00
lyingbug 877a6ebf9d feat(api): add resource_urls=public to return directly loadable file URLs
API responses reference stored files as opaque resource:// handles, so an
integrating app had to make a second authenticated call to /files for every
image before it could render anything.

Add an opt-in that resolves those references server-side into time-limited
HTTP(S) URLs, using the same mechanism the IM channels already rely on:

  - per request: ?resource_urls=public (default: handle, unchanged)
  - per deployment: RESOURCE_URL_MODE=public

Applied to the chat SSE endpoints (knowledge-chat, agent-chat,
continue-stream), message history load, and knowledge-search. Streamed answers
buffer a trailing incomplete reference so a handle split across two deltas is
still rewritten. References that cannot become an HTTP URL (for example local
storage with no APP_EXTERNAL_URL) stay handles, so clients keep the /files
fallback. Embed channels are deliberately excluded: their visitors are
anonymous.

SSE payloads share SearchResult pointers and metadata maps with the stream
replay buffer and the message being persisted, so those are rewritten as copies.

Co-authored-by: lyingbug <lyingbug@users.noreply.github.com>
2026-08-05 14:42:51 +08:00
lyingbug a6926fc912 refactor(storage): extract storage reference URL rewriting into internal/storageurl
The IM channels rewrite resource:// and provider:// references into loadable
HTTP URLs because IM clients cannot attach WeKnora credentials to an image
fetch. That logic was private to internal/im, so no other surface could reuse
it.

Move the reference pattern, HTTP-result guard, tenant-aware storage backend
resolver and streaming holdback helpers into internal/storageurl, and add a
Rewriter that memoises resolutions (a resource:// handle costs one access-grant
row per resolution). internal/im now delegates; behaviour is unchanged.

Co-authored-by: lyingbug <lyingbug@users.noreply.github.com>
2026-08-05 14:42:51 +08:00
xiaowang 51b88bbae3 fix(summary): preserve rows across table chunks / 修复表格分块摘要丢行 2026-08-05 13:12:59 +08:00
qilifan 3507156929
feat(storage): support AWS S3 NoAk default credentials (#2008)
* feat(storage): support AWS S3 default credentials

* fix(storage): allow clearing S3 credentials

* fix static checks and trim locale diff

---------

Co-authored-by: qilifan <4359902+qilifan@users.noreply.github.com>
2026-08-05 11:59:02 +08:00
wizardchen 350d9e4eca perf(container): batch legacy storage alias lookup at startup
The startup migration probed every workspace/provider pair with First(),
so each miss emitted a GORM "record not found" log line and a separate
round trip. Load all aliases once and index them in memory instead.
2026-08-04 22:48:54 +08:00
wizardchen 5723e837f8 fix(chat): cover web partial success and wiki-only embed fallback
Refine embed-failure target filtering so wiki/graph-only KBs and missing
KB metadata still reach HybridSearch instead of hard search_failed.
Add regression tests for KB failure with concurrent web hits and wiki-only
empty recall after embedding errors.
2026-08-04 21:12:51 +08:00
KKtwo 22af9b2f78 fix(chat): preserve retrieval fallback errors 2026-08-04 21:12:51 +08:00
wizardchen d37e289309 fix(frontend): restore runtime queue i18n and guard against prune
Re-add worker pool, queue, and task type labels dropped during locale
cleanup, register dynamic key prefixes in the i18n audit, and add tests
so check-i18n fails if they are removed again.
2026-08-04 21:07:26 +08:00
wizardchen 2f2be9f601 refactor(knowledge): improve folder path handling and add tests for underscore scenarios
- Updated SQL queries to properly escape LIKE wildcards in folder path filters.
- Added tests for renaming knowledge folder paths, including cases with underscores in paths.
- Enhanced existing tests to ensure correct folder listing behavior when using underscores.
2026-08-04 20:45:29 +08:00
wizardchen bb23444e39 fix(knowledge): sync folder APIs and address review follow-ups
Document folder tree/move/rename endpoints in docs/api, Swagger, and the Go
client; align frontend path normalization with server caps; guard rename
no-ops and duplicate folder creation; enforce KB ownership on folder rename.
2026-08-04 20:45:29 +08:00
wizardchen 7316d95f56 feat(knowledge): enhance folder management and upload confirmation dialog
- Added folder options to the upload confirmation dialog, allowing users to select from existing folders as upload destinations.
- Updated folder tree UI to improve folder browsing experience, displaying sub-folders and documents more intuitively.
- Enhanced folder picker functionality to support creating new sub-folders directly from the picker.
- Improved localization for folder-related strings across multiple languages.
- Refactored related components and tests to accommodate new features and ensure proper functionality.

This commit streamlines the user experience for managing folders and uploading documents, making it easier to navigate and organize content.
2026-08-04 20:45:29 +08:00
lyingbug 192a721061 fix(knowledge): stop the root row rendering a rename input, and demote the folder picker to a popup
Two things the folder UI got wrong.

The rename state used the empty string to mean "nothing is being renamed",
but the empty string is also the root folder's own path, so the root row
matched and permanently rendered a stray rename input. The sentinel is now
null, and rename is additionally gated on the row being a real folder, so no
sentinel value can reach the root again.

Moving documents also did not deserve a modal dialog: it is one small,
reversible choice. In a row's menu the picker is now another level of the menu
that is already open, mirroring the "move to knowledge base" sub-menu right
next to it; in the batch bar it hangs off the button. Picking a folder performs
the move, since there is nothing else to confirm.

The upload destination line was also mangling folder names: truncating from the
left with direction: rtl reorders CJK text. It now shows the destination's last
segment with the full path in the title, reads as a quiet line instead of a
grey banner, and only appears when the destination is not the top level.
2026-08-04 20:45:29 +08:00
lyingbug 5fea7a1695 feat(knowledge): browse folders like a file manager and allow re-filing
The scope switch is gone. It existed because "select a folder" was ambiguous
in a list that could not show folders, and turning that ambiguity into a
control just moved the problem onto the user.

The list now shows what a folder contains: its sub-folders as entries first,
then the documents directly inside it. So the mode follows what the user is
already doing instead of a switch - browsing shows one level, and any active
filter searches the whole subtree flat, with a breadcrumb note saying so. A
document uploaded on its own needs no special row either: it sits at the top
level next to the folders, exactly where it is.

Folders can also be adjusted now, which they could not before - a mis-filed
upload previously had to be deleted and re-uploaded. Documents move via the
row menu or the batch bar into an existing or newly typed folder, and a
folder can be renamed in place, carrying its subtree. Both are plain
folder_path updates: nothing is re-parsed, re-chunked or re-embedded.
2026-08-04 20:45:29 +08:00
lyingbug b3e31a24d9 fix(knowledge): give individually uploaded documents a visible home in the tree
A folder-only tree cannot show a node for a document that is not in any
folder, so single-file uploads were unreachable from the sidebar: a base
with 5 documents and one 3-document folder just left two of them
unaccounted for.

Rather than add a second pseudo-folder row for them (the ambiguity the
previous commit removed), the one existing mechanism is made visible: the
scope switch is now a labeled segmented control instead of an icon behind a
tooltip, and the root row is named after whatever the current scope makes it
list - the whole base, or exactly the unfiled documents. The breadcrumb's
leading crumb follows the same rule.
2026-08-04 20:45:29 +08:00
lyingbug d6083c6c98 fix(knowledge): make the folder tree root a real parent node
The sidebar listed "all documents" and "root" as two rows sitting next to
the top-level folders, but those folders are inside the root, so the three
rows contradicted each other and reaching the top level for an upload meant
selecting a row that did not look like a parent.

There is now a single root row with every folder nested beneath it, and what
a row lists is decided by one uniform rule at every level: the direct
documents, or the whole subtree when the scope toggle is on. "Documents not
in any folder" is the root with the toggle off, which is also why the toggle
now defaults to on.

The upload destination is no longer implied by the sidebar selection either:
the confirmation dialog shows which folder the batch lands in, lets it be
switched back to the top level, and lists each file's own sub-directory.
2026-08-04 20:45:29 +08:00
lyingbug 27efebbd1c refactor(knowledge): extract folder tree helpers and unit-test them
The upload-path builder, breadcrumb/ancestor derivation, folder lookup and
tree flattening were inline in the view and the sidebar component. Moving
them into folderTree.ts follows the convention of kbListMerge.ts and
wikiStatusRefresh.ts, removes the duplicated webkitRelativePath handling,
and makes the folder navigation logic directly testable.
2026-08-04 20:45:29 +08:00
lyingbug f747d390a8 feat(knowledge): show uploaded folder structure as a sidebar tree
The documents tab gains a left folder column built from the folder_path
now stored on each knowledge entry. Selecting a folder filters the
existing document list (so search, tags, type, status and date filters
all keep working), a breadcrumb above the list shows where you are, and
uploads made while browsing a folder land in that folder. A per-row
folder chip appears only when the list can span folders.
2026-08-04 20:45:29 +08:00
lyingbug fa67dfd44b feat(knowledge): store folder path for folder uploads and expose folder tree API
Folder uploads used to encode their relative directory inside file_name,
which made the document list render the whole path as the title and left
no way to browse or query a single folder. The directory now lives in a
dedicated folder_path column (with a backfill for existing rows), the
document list accepts folder_path / folder_recursive filters, and a new
GET /knowledge-bases/:id/knowledge/folders endpoint returns the folder
tree with per-folder document counts.
2026-08-04 20:45:29 +08:00
ljl 2b31d3681b fix(embedding): prevent ants worker panic and mutex deadlock on empty/mismatched results
If the embedding provider returns an empty slice of vectors (e.g. length 0)
due to api mismatch or empty input, BatchEmbed will return length 0.
Accessing embedding[i] directly in BatchEmbedWithPool results in an out of bounds panic.
Because the panic occurs inside a mu.Lock() block, the mutex is never unlocked,
leading to a permanent deadlock of the document processing queue.

This fix checks the returned embeddings slice length and returns an error
instead of panicking, releasing the mutex properly.
2026-08-04 16:46:57 +08:00
lyingbug fcc4cd6a9f fix(frontend): announce RAG wait status from a persistent live region
role="status" lived on the wait row itself, so the live region was inserted
together with its own text and screen readers had nothing to announce. Move the
announcement to an sr-only region that stays mounted for the whole turn, which
also covers the pre-pipeline wait row that had no announcement at all.

Co-authored-by: lyingbug <lyingbug@users.noreply.github.com>
2026-08-04 13:21:23 +08:00
lyingbug 2454eaf253 fix(frontend): cover attachment-only turns and stalled streams in the RAG wait row
The wait row only appeared once a completed knowledge_search step existed, so
attachment-only turns (attachment_parsing / image_analysis with no KB hit) still
showed nothing between the last step and the first answer token. Model waiting is
now a three-state kind: 'model' after retrieval finished, the neutral 'preparing'
row otherwise.

A dropped SSE connection never sets is_completed (the stream layer only raises a
toast), so the row used to promise an answer forever. It now stops claiming
progress after RAG_WAIT_STALL_DELAY_MS and drops the shimmer.

Also share RAG_RETRIEVAL_TOOL_NAMES instead of repeating the knowledge_search /
search_knowledge alias check, name the reveal delay, and move the reveal/stall
timers into a controller with an injectable scheduler so the timing is covered by
tests instead of by source-matching alone.

Co-authored-by: lyingbug <lyingbug@users.noreply.github.com>
2026-08-04 13:21:23 +08:00
ttommybot 747cbcbc02 fix: show answer generation status after retrieval 2026-08-04 13:21:23 +08:00
ttommybot 91b2d31cd4 fix: unify organization settings modal scrolling 2026-08-03 23:54:10 +08:00
lyingbug 82ea91f18e fix(knowledge): clear stale error_message on the remaining reprocess paths
Follow-up to the error_message clearing added for SetFinalizing,
FinalizeSubtask and ProcessDocument. Two routes still persisted a
previous attempt's failure message:

- ProcessManualUpdate flips a row back to 'processing' without clearing
  error_message. A manual knowledge item that failed earlier (for
  example when the knowledge base lookup failed, or after an Asynq
  retry) re-enters processing still carrying that message, so the UI
  shows a failure on a document it also shows as running.

- finalizeIndexedKnowledgeState is the only route to 'completed' that
  bypasses FinalizeSubtask. When a document has no text chunks and no
  pending multimodal work it is completed in place, so the promotion
  that would otherwise clear the column never runs and a successfully
  indexed row keeps reporting the earlier failure.

The 'processing' transition is extracted into markKnowledgeProcessing so
ProcessDocument and ProcessManualUpdate share one definition of what
starting a new attempt means, and a future worker cannot reintroduce the
same omission. The helper also replaces the "processing" string literal
with types.ParseStatusProcessing.

Tests: markKnowledgeProcessing gets direct coverage, and the
finalizeIndexedKnowledgeState table now seeds a stale error_message and
asserts it survives the processing branches (FinalizeSubtask clears it
later) and is cleared on the completed branch.
2026-08-03 20:14:09 +08:00
shangfeng-liu 44206e4ad4 fix(knowledge): clear stale error_message on reprocessing and finalize
A knowledge item that failed once keeps error_message set; when it is
reprocessed or driven through the finalizing pipeline, the stale message
was never cleared, so the UI kept showing an outdated failure on items
that were processing or had completed successfully.

- SetFinalizing and FinalizeSubtask now clear error_message in the same
  atomic UPDATE that transitions parse_status
- ProcessDocument clears knowledge.ErrorMessage when flipping the row
  to processing for a new attempt

Add a SQLite-backed regression test covering both transitions.
2026-08-03 19:38:22 +08:00
shiquan_zhan ccb0cee3e6 fix: exclude archived wiki pages from lint cursor 2026-08-03 17:30:41 +08:00
wizardchen a8cf3a2b09 fix(mcp-server): normalize shared KB list shape for name resolution
GET /shared-knowledge-bases returns rows with nested knowledge_base
(sharedKBRow), unlike GET /knowledge-bases which returns flat KB objects.
resolve_kb_id and agent_chat preflight were reading top-level id/name,
so shared KB names never matched. Add _normalize_kb_entries and tests.
2026-08-03 15:50:33 +08:00
FincherYuan 1e75a34300 test(mcp-server): bump tool count assertion to 29 for new shared-KB tool
Co-Authored-By: Claude <noreply@anthropic.com>
2026-08-03 15:50:33 +08:00
FincherYuan 71bdecf873 fix(mcp-server): integrate shared KBs into name resolution and tool hints
resolve_kb_id now searches both own and shared knowledge bases when
resolving a name to UUID. Error messages and docstrings in hybrid_search,
chat, and agent_chat now reference list_shared_knowledge_bases as well.

Co-Authored-By: Claude <noreply@anthropic.com>
2026-08-03 15:50:33 +08:00
FincherYuan 5f4464db8d feat(mcp-server): add list_shared_knowledge_bases tool
Add a new MCP tool to list knowledge bases shared from other workspaces
via the GET /shared-knowledge-bases endpoint, which was previously not
exposed through the MCP server.

- Add WeKnoraClient.list_shared_knowledge_bases() method
- Register list_shared_knowledge_bases() as an @mcp.tool()

Co-Authored-By: Claude <noreply@anthropic.com>
2026-08-03 15:50:33 +08:00
shangfeng-liu 583b87b1b9 fix: preserve MinerU automatic PDF parsing 2026-08-03 15:37:07 +08:00
ttommybot d3055ae068 fix: count skipped stages in parsing progress 2026-08-03 15:20:16 +08:00
xiaowang 60028cbe0e fix(rerank): preserve code and math passages 2026-08-03 15:10:09 +08:00
xiaowang 31a2358bbf fix(rerank): normalize NVIDIA logit scores 2026-08-03 15:07:15 +08:00