test(rerank): pin every vendor's request, and correct three facts the table surfaced (#3513)
parent
cc82e75cb5
commit
4640a1ea0c
|
|
@ -67,7 +67,6 @@ func TestRerankProtocolAssignment(t *testing.T) {
|
|||
"qianfan": api.RerankCohere,
|
||||
"gpustack": api.RerankCohere,
|
||||
"generic": api.RerankCohere,
|
||||
"openai": api.RerankCohere,
|
||||
"weknoracloud": api.RerankCohere,
|
||||
"novita": api.RerankCohere,
|
||||
"openrouter": api.RerankCohere,
|
||||
|
|
@ -263,8 +262,8 @@ func TestRerankCeilingsAreTheDocumentedOnes(t *testing.T) {
|
|||
// cloud.tencent.com/document/product/1772: RunRerank 60 docs,
|
||||
// Query + Docs together 2000 characters, one request at a time.
|
||||
"lkeap": {MaxDocuments: 60, MaxRequestChars: 2000, MaxConcurrency: 1},
|
||||
// Ark Knowledge Service rerank: 50 per request.
|
||||
"volcengine": {MaxDocuments: 50, MaxConcurrency: 4},
|
||||
// VikingDB Knowledge Service rerank: datas "数组长度不超过 200".
|
||||
"volcengine": {MaxDocuments: 200, MaxConcurrency: 4},
|
||||
// NIM reranking: passages is capped at 512 items.
|
||||
"nvidia": {MaxDocuments: 512},
|
||||
// cloud.baidu.com/doc/qianfan-api: 文本数量不超过64, query 不超过
|
||||
|
|
|
|||
|
|
@ -21,6 +21,14 @@ const (
|
|||
LKEAPDefaultRerankModel = "lke-reranker-base"
|
||||
)
|
||||
|
||||
// lkeapEndpoint and lkeapScheme are where the SDK sends RunRerank: the public
|
||||
// API host, as it has always been, whatever the row's base URL says. They are
|
||||
// variables only so a test can point the SDK at a local server.
|
||||
var (
|
||||
lkeapEndpoint = LKEAPRerankEndpoint
|
||||
lkeapScheme = "HTTPS"
|
||||
)
|
||||
|
||||
// lkeapClient calls Tencent Cloud's RunRerank action through the official
|
||||
// SDK, which owns the TC3 signing this API requires. It implements
|
||||
// api.Reranker and nothing else: the per-request ceilings (60 documents,
|
||||
|
|
@ -50,7 +58,8 @@ func newLKEAPClient(config *RerankerConfig, resolved *catalog.Resolved) (api.Rer
|
|||
|
||||
credential := common.NewCredential(secretID, secretKey)
|
||||
cpf := profile.NewClientProfile()
|
||||
cpf.HttpProfile.Endpoint = LKEAPRerankEndpoint
|
||||
cpf.HttpProfile.Endpoint = lkeapEndpoint
|
||||
cpf.HttpProfile.Scheme = lkeapScheme
|
||||
|
||||
client, err := lkeap.NewClient(credential, region, cpf)
|
||||
if err != nil {
|
||||
|
|
|
|||
|
|
@ -65,7 +65,10 @@ func TestVolcengineClientOutboundShape(t *testing.T) {
|
|||
require.NotNil(t, request.Datas[0].Content)
|
||||
assert.Equal(t, "会保留", *request.Datas[0].Content)
|
||||
require.NotNil(t, request.RerankInstruction)
|
||||
assert.Contains(t, *request.RerankInstruction, "Document")
|
||||
// The console's default, verbatim, so results match what the console shows.
|
||||
assert.Equal(t, volcengineRerankDefaultInstruction, *request.RerankInstruction)
|
||||
assert.Equal(t, "Whether the document answers the query or matches the content retrieval intent",
|
||||
*request.RerankInstruction)
|
||||
}
|
||||
|
||||
func TestVolcengineClientSurfacesAnInBodyErrorCode(t *testing.T) {
|
||||
|
|
|
|||
|
|
@ -13,14 +13,17 @@ import (
|
|||
)
|
||||
|
||||
const (
|
||||
volcengineRerankDefaultModel = "doubao-seed-rerank"
|
||||
volcengineRerankDefaultRegion = "cn-beijing"
|
||||
volcengineRerankDefaultInstruction = "Whether the Document answers the Query or matches the content retrieval intent"
|
||||
volcengineRerankDefaultModel = "doubao-seed-rerank"
|
||||
volcengineRerankDefaultRegion = "cn-beijing"
|
||||
// The console's default instruction, verbatim: "如需对齐控制台效果,请使用
|
||||
// 相同指令" (https://docs.volcengine.com/docs/vector_database_vikingdb/Rerank).
|
||||
volcengineRerankDefaultInstruction = "Whether the document answers the query " +
|
||||
"or matches the content retrieval intent"
|
||||
)
|
||||
|
||||
// volcengineClient calls the managed Knowledge Service rerank through the
|
||||
// vikingdb SDK, which owns the AK/SK signing. It implements api.Reranker and
|
||||
// nothing else: the 50-document ceiling and the batch concurrency are
|
||||
// nothing else: the 200-document ceiling and the batch concurrency are
|
||||
// declared on the vendor and enforced by protocolReranker.
|
||||
type volcengineClient struct {
|
||||
modelName string
|
||||
|
|
|
|||
|
|
@ -0,0 +1,357 @@
|
|||
package rerank
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"fmt"
|
||||
"math"
|
||||
"net/http"
|
||||
"net/http/httptest"
|
||||
"sort"
|
||||
"strings"
|
||||
"sync"
|
||||
"testing"
|
||||
"unicode/utf8"
|
||||
|
||||
"github.com/Tencent/WeKnora/internal/models/catalog"
|
||||
"github.com/Tencent/WeKnora/internal/types"
|
||||
"github.com/stretchr/testify/assert"
|
||||
"github.com/stretchr/testify/require"
|
||||
)
|
||||
|
||||
// recorded is one request as the upstream saw it.
|
||||
type recorded struct {
|
||||
path string
|
||||
header http.Header
|
||||
body map[string]any
|
||||
}
|
||||
|
||||
// upstream is a stand-in for every rerank vendor at once. It answers in
|
||||
// whichever shape the request belongs to, scoring each document by its
|
||||
// length so that the expected ranking is known: longer is more relevant.
|
||||
// Probability vendors get length/100, logit vendors length-3, which puts the
|
||||
// shortest documents below zero the way a real NIM reply does.
|
||||
type upstream struct {
|
||||
mu sync.Mutex
|
||||
requests []recorded
|
||||
url string
|
||||
}
|
||||
|
||||
func newUpstream(t *testing.T) *upstream {
|
||||
t.Helper()
|
||||
withRerankSSRFWhitelist(t, "127.0.0.1")
|
||||
u := &upstream{}
|
||||
server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
|
||||
var body map[string]any
|
||||
if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
|
||||
http.Error(w, err.Error(), http.StatusBadRequest)
|
||||
return
|
||||
}
|
||||
u.mu.Lock()
|
||||
u.requests = append(u.requests, recorded{r.URL.Path, r.Header.Clone(), body})
|
||||
u.mu.Unlock()
|
||||
w.Header().Set("Content-Type", "application/json")
|
||||
_, _ = w.Write([]byte(answer(r, body)))
|
||||
}))
|
||||
t.Cleanup(server.Close)
|
||||
u.url = server.URL
|
||||
|
||||
// LKEAP's SDK always calls the public API host; point it here.
|
||||
savedEndpoint, savedScheme := lkeapEndpoint, lkeapScheme
|
||||
lkeapEndpoint, lkeapScheme = strings.TrimPrefix(server.URL, "http://"), "HTTP"
|
||||
t.Cleanup(func() { lkeapEndpoint, lkeapScheme = savedEndpoint, savedScheme })
|
||||
return u
|
||||
}
|
||||
|
||||
// texts pulls the documents out of whichever shape carried them.
|
||||
func texts(v any) []string {
|
||||
var out []string
|
||||
for _, item := range v.([]any) {
|
||||
switch d := item.(type) {
|
||||
case string:
|
||||
out = append(out, d)
|
||||
case map[string]any: // NIM passages, Volcengine datas
|
||||
if text, ok := d["text"].(string); ok {
|
||||
out = append(out, text)
|
||||
} else {
|
||||
out = append(out, d["content"].(string))
|
||||
}
|
||||
}
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
func probability(doc string) float64 { return float64(utf8.RuneCountInString(doc)) / 100 }
|
||||
func logit(doc string) float64 { return float64(utf8.RuneCountInString(doc)) - 3 }
|
||||
|
||||
// answer replies in the protocol the request belongs to. The indexed shapes
|
||||
// come back in reverse order so that placement by index is exercised.
|
||||
func answer(r *http.Request, body map[string]any) string {
|
||||
var parts []string
|
||||
switch {
|
||||
case r.Header.Get("X-TC-Action") == "RunRerank": // LKEAP: scores in input order
|
||||
for _, doc := range texts(body["Docs"]) {
|
||||
parts = append(parts, fmt.Sprint(probability(doc)))
|
||||
}
|
||||
return `{"Response":{"ScoreList":[` + strings.Join(parts, ",") + `],"RequestId":"r"}}`
|
||||
case strings.HasSuffix(r.URL.Path, "/api/knowledge/service/rerank"): // Volcengine
|
||||
for _, doc := range texts(body["datas"]) {
|
||||
parts = append(parts, fmt.Sprint(probability(doc)))
|
||||
}
|
||||
return `{"code":0,"message":"success","data":{"scores":[` + strings.Join(parts, ",") + `]}}`
|
||||
case strings.HasSuffix(r.URL.Path, "/reranking"): // NIM
|
||||
docs := texts(body["passages"])
|
||||
for i := len(docs) - 1; i >= 0; i-- {
|
||||
parts = append(parts, fmt.Sprintf(`{"index":%d,"logit":%v}`, i, logit(docs[i])))
|
||||
}
|
||||
return `{"rankings":[` + strings.Join(parts, ",") + `]}`
|
||||
case strings.Contains(r.URL.Path, "/text-rerank"): // DashScope
|
||||
docs := texts(body["input"].(map[string]any)["documents"])
|
||||
for i := len(docs) - 1; i >= 0; i-- {
|
||||
parts = append(parts, fmt.Sprintf(`{"index":%d,"relevance_score":%v}`, i, probability(docs[i])))
|
||||
}
|
||||
return `{"output":{"results":[` + strings.Join(parts, ",") + `]}}`
|
||||
default: // Cohere
|
||||
docs := texts(body["documents"])
|
||||
score := probability
|
||||
if r.Header.Get("X-Test-Scale") == "logit" {
|
||||
score = logit
|
||||
}
|
||||
for i := len(docs) - 1; i >= 0; i-- {
|
||||
parts = append(parts, fmt.Sprintf(`{"index":%d,"relevance_score":%v}`, i, score(docs[i])))
|
||||
}
|
||||
return `{"results":[` + strings.Join(parts, ",") + `]}`
|
||||
}
|
||||
}
|
||||
|
||||
func documents(n int) []string {
|
||||
out := make([]string, n)
|
||||
for i := range out {
|
||||
out[i] = strings.Repeat("d", i%7+1) + fmt.Sprint(i)
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
func anyStrings(in []string) []any {
|
||||
out := make([]any, len(in))
|
||||
for i, s := range in {
|
||||
out[i] = s
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
// TestRerankWireFormatPerVendor pins, for every rerank vendor in the catalog,
|
||||
// the request the factory actually sends — URL, credential and the complete
|
||||
// body — and that the answer comes back ranked on the 0..1 scale the
|
||||
// retrieval threshold assumes. Each expectation traces to the vendor's
|
||||
// reference as cited in its vendor.go.
|
||||
func TestRerankWireFormatPerVendor(t *testing.T) {
|
||||
const query = "q"
|
||||
three := []string{"a", "bbb", "cc"}
|
||||
cohere := func(model string, docs []string, extra map[string]any) map[string]any {
|
||||
body := map[string]any{"model": model, "query": query, "documents": anyStrings(docs)}
|
||||
for k, v := range extra {
|
||||
body[k] = v
|
||||
}
|
||||
return body
|
||||
}
|
||||
const volcInstruction = "Whether the document answers the query or matches the content retrieval intent"
|
||||
|
||||
cases := []struct {
|
||||
name string
|
||||
provider string
|
||||
model string
|
||||
base string // appended to the upstream URL
|
||||
extra map[string]string
|
||||
appID string
|
||||
appSecret string
|
||||
docs []string
|
||||
logitScale bool // the protocol answers raw logits (NIM)
|
||||
// cohereLogs marks a Cohere-shape vendor that declares a logit scale
|
||||
// (GPUStack). The shape cannot tell the stub which scale to answer
|
||||
// in, so the case asks for logits through a test-only header.
|
||||
cohereLogs bool
|
||||
|
||||
wantPath string
|
||||
wantAuth [2]string // header, a substring its value must contain
|
||||
wantRequests int
|
||||
// wantBody is one request's body, exactly. Batches go out
|
||||
// concurrently, so it is matched against every request rather than
|
||||
// the first to arrive.
|
||||
wantBody map[string]any
|
||||
}{
|
||||
{
|
||||
name: "jina echoes documents", provider: "jina", model: "jina-reranker-v3", base: "/v1",
|
||||
wantPath: "/v1/rerank", wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: cohere("jina-reranker-v3", three, map[string]any{"return_documents": true}),
|
||||
},
|
||||
{
|
||||
name: "zhipu on its full rerank URL", provider: "zhipu", model: "rerank", base: "/api/paas/v4/rerank",
|
||||
wantPath: "/api/paas/v4/rerank", wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: cohere("rerank", three, map[string]any{"return_documents": true}),
|
||||
},
|
||||
{
|
||||
name: "zhipu splits at 128 documents", provider: "zhipu", model: "rerank", base: "/api/paas/v4/rerank",
|
||||
docs: documents(129), wantRequests: 2,
|
||||
wantPath: "/api/paas/v4/rerank", wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: cohere("rerank", documents(128), map[string]any{"return_documents": true}),
|
||||
},
|
||||
{
|
||||
name: "siliconflow", provider: "siliconflow", model: "BAAI/bge-reranker-v2-m3", base: "/v1",
|
||||
wantPath: "/v1/rerank", wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: cohere("BAAI/bge-reranker-v2-m3", three, nil),
|
||||
},
|
||||
{
|
||||
name: "qianfan splits at 64 documents", provider: "qianfan", model: "bce-reranker-base", base: "/v2",
|
||||
docs: documents(65), wantRequests: 2,
|
||||
wantPath: "/v2/rerank", wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: cohere("bce-reranker-base", documents(64), nil),
|
||||
},
|
||||
{
|
||||
name: "gpustack sends the row's truncation budget and answers logits", provider: "gpustack",
|
||||
model: "bge-reranker-v2-m3", base: "/v1",
|
||||
extra: map[string]string{catalog.ExtraTruncatePromptTokens: "256"},
|
||||
cohereLogs: true,
|
||||
wantPath: "/v1/rerank", wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: cohere("bge-reranker-v2-m3", three, map[string]any{"truncate_prompt_tokens": float64(256)}),
|
||||
},
|
||||
{
|
||||
name: "generic", provider: "generic", model: "bge-reranker-v2-m3", base: "/v1",
|
||||
wantPath: "/v1/rerank", wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: cohere("bge-reranker-v2-m3", three, nil),
|
||||
},
|
||||
{
|
||||
name: "novita", provider: "novita", model: "baai/bge-reranker-v2-m3", base: "/openai/v1",
|
||||
wantPath: "/openai/v1/rerank", wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: cohere("baai/bge-reranker-v2-m3", three, nil),
|
||||
},
|
||||
{
|
||||
name: "openrouter", provider: "openrouter", model: "cohere/rerank-v3.5", base: "/api/v1",
|
||||
wantPath: "/api/v1/rerank", wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: cohere("cohere/rerank-v3.5", three, nil),
|
||||
},
|
||||
{
|
||||
name: "litellm on the proxy root", provider: "litellm", model: "cohere/rerank-v3.5",
|
||||
wantPath: "/rerank", wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: cohere("cohere/rerank-v3.5", three, nil),
|
||||
},
|
||||
{
|
||||
name: "weknoracloud signs its own path", provider: "weknoracloud", model: "rerank",
|
||||
appID: "app", appSecret: "s",
|
||||
wantPath: "/api/v1/rerank", wantAuth: [2]string{"X-APPID", "app"},
|
||||
wantBody: cohere("rerank", three, nil),
|
||||
},
|
||||
{
|
||||
name: "aliyun wraps input and parameters", provider: "aliyun", model: "gte-rerank-v2",
|
||||
base: "/api/v1/services/rerank/text-rerank/text-rerank",
|
||||
wantPath: "/api/v1/services/rerank/text-rerank/text-rerank",
|
||||
wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: map[string]any{
|
||||
"model": "gte-rerank-v2",
|
||||
"input": map[string]any{"query": query, "documents": anyStrings(three)},
|
||||
"parameters": map[string]any{"return_documents": true, "top_n": float64(3)},
|
||||
},
|
||||
},
|
||||
{
|
||||
name: "nvidia sends passages and answers logits", provider: "nvidia",
|
||||
model: "nvidia/nv-rerankqa-mistral-4b-v3", base: "/v1/retrieval/nvidia/reranking", logitScale: true,
|
||||
wantPath: "/v1/retrieval/nvidia/reranking", wantAuth: [2]string{"Authorization", "Bearer k"},
|
||||
wantBody: map[string]any{
|
||||
"model": "nvidia/nv-rerankqa-mistral-4b-v3",
|
||||
"query": map[string]any{"text": query},
|
||||
"passages": []any{
|
||||
map[string]any{"text": "a"}, map[string]any{"text": "bbb"}, map[string]any{"text": "cc"},
|
||||
},
|
||||
"truncate": "END",
|
||||
},
|
||||
},
|
||||
{
|
||||
name: "volcengine through the Knowledge Service SDK", provider: "volcengine", model: "doubao-seed-rerank",
|
||||
appSecret: "secret-test",
|
||||
wantPath: "/api/knowledge/service/rerank", wantAuth: [2]string{"Authorization", "Credential=k/"},
|
||||
wantBody: map[string]any{
|
||||
"rerank_model": "doubao-seed-rerank",
|
||||
"rerank_instruction": volcInstruction,
|
||||
"datas": []any{
|
||||
map[string]any{"query": query, "content": "a"},
|
||||
map[string]any{"query": query, "content": "bbb"},
|
||||
map[string]any{"query": query, "content": "cc"},
|
||||
},
|
||||
},
|
||||
},
|
||||
{
|
||||
name: "lkeap through the Tencent Cloud SDK", provider: "lkeap", model: "lke-reranker-base",
|
||||
appSecret: "secret-test",
|
||||
wantPath: "/", wantAuth: [2]string{"Authorization", "Credential=k/"},
|
||||
wantBody: map[string]any{"Query": query, "Docs": anyStrings(three), "Model": "lke-reranker-base"},
|
||||
},
|
||||
}
|
||||
|
||||
for _, tc := range cases {
|
||||
t.Run(tc.name, func(t *testing.T) {
|
||||
up := newUpstream(t)
|
||||
docs := tc.docs
|
||||
if docs == nil {
|
||||
docs = three
|
||||
}
|
||||
headers := map[string]string{}
|
||||
if tc.cohereLogs {
|
||||
headers["X-Test-Scale"] = "logit"
|
||||
}
|
||||
reranker, err := NewReranker(&RerankerConfig{
|
||||
Source: types.ModelSourceRemote, Provider: tc.provider, ModelName: tc.model,
|
||||
BaseURL: up.url + tc.base, APIKey: "k", ExtraConfig: tc.extra,
|
||||
AppID: tc.appID, AppSecret: tc.appSecret, CustomHeaders: headers,
|
||||
})
|
||||
require.NoError(t, err)
|
||||
|
||||
got, err := reranker.Rerank(context.Background(), query, docs)
|
||||
require.NoError(t, err)
|
||||
|
||||
wantRequests := tc.wantRequests
|
||||
if wantRequests == 0 {
|
||||
wantRequests = 1
|
||||
}
|
||||
require.Len(t, up.requests, wantRequests)
|
||||
bodies := make([]map[string]any, 0, len(up.requests))
|
||||
for _, req := range up.requests {
|
||||
assert.Equal(t, tc.wantPath, req.path)
|
||||
assert.Contains(t, req.header.Get(tc.wantAuth[0]), tc.wantAuth[1])
|
||||
if tc.appSecret != "" {
|
||||
assert.NotContains(t, req.header.Get("Authorization"), tc.appSecret,
|
||||
"a secret key signs the request and must never travel")
|
||||
}
|
||||
bodies = append(bodies, req.body)
|
||||
}
|
||||
assert.Contains(t, bodies, tc.wantBody)
|
||||
|
||||
// Every document comes back once, ranked by the upstream's score,
|
||||
// and on the probability scale.
|
||||
require.Len(t, got, len(docs))
|
||||
want := append([]string(nil), docs...)
|
||||
sort.SliceStable(want, func(i, j int) bool {
|
||||
return utf8.RuneCountInString(want[i]) > utf8.RuneCountInString(want[j])
|
||||
})
|
||||
for i, r := range got {
|
||||
assert.Equal(t, docs[r.Index], r.Document.Text, "result %d carries the wrong text", i)
|
||||
assert.Equal(t, utf8.RuneCountInString(want[i]), utf8.RuneCountInString(r.Document.Text),
|
||||
"result %d is out of rank", i)
|
||||
assert.GreaterOrEqual(t, r.RelevanceScore, 0.0)
|
||||
assert.LessOrEqual(t, r.RelevanceScore, 1.0)
|
||||
}
|
||||
if tc.logitScale || tc.cohereLogs {
|
||||
// length 1 is a logit of -2: well below one half once converted.
|
||||
last := got[len(got)-1]
|
||||
assert.InDelta(t, 1/(1+math.Exp(2)), last.RelevanceScore, 1e-9,
|
||||
"a logit must be converted, not passed through")
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
// OpenAI has no rerank API, so the vendor does not offer the type.
|
||||
func TestOpenAIDoesNotOfferRerank(t *testing.T) {
|
||||
v, ok := catalog.Get("openai")
|
||||
require.True(t, ok)
|
||||
assert.False(t, v.SupportsType(types.ModelTypeRerank))
|
||||
}
|
||||
|
|
@ -21,8 +21,8 @@
|
|||
// gpt-5* / o1* / o3* / o4* families set supports_temperature false;
|
||||
// - `store: false` is sent so completions are not retained, and
|
||||
// `prompt_cache_key` steers cache routing; usage reports cached tokens;
|
||||
// - every model type (chat, embedding, rerank via generic clients, VLM,
|
||||
// ASR) is served from https://api.openai.com/v1;
|
||||
// - chat, embedding, VLM and ASR are served from https://api.openai.com/v1;
|
||||
// there is no rerank API;
|
||||
// - requests that go to api.openai.com itself use the Responses protocol
|
||||
// (https://developers.openai.com/api/docs/api-reference/responses),
|
||||
// OpenAI's primary API where reasoning items and encrypted reasoning are
|
||||
|
|
@ -85,14 +85,16 @@ func init() {
|
|||
DefaultBaseURLs: map[types.ModelType]string{
|
||||
types.ModelTypeKnowledgeQA: BaseURL,
|
||||
types.ModelTypeEmbedding: BaseURL,
|
||||
types.ModelTypeRerank: BaseURL,
|
||||
types.ModelTypeVLLM: BaseURL,
|
||||
types.ModelTypeASR: BaseURL,
|
||||
},
|
||||
// No rerank: OpenAI's API has no rerank endpoint — the reference
|
||||
// index (https://developers.openai.com/api/llms.txt) does not mention
|
||||
// one — so a row created here would only ever 404. A relay that
|
||||
// serves rerank behind an OpenAI-style URL is a generic row.
|
||||
ModelTypes: []types.ModelType{
|
||||
types.ModelTypeKnowledgeQA,
|
||||
types.ModelTypeEmbedding,
|
||||
types.ModelTypeRerank,
|
||||
types.ModelTypeVLLM,
|
||||
types.ModelTypeASR,
|
||||
},
|
||||
|
|
|
|||
|
|
@ -22,8 +22,8 @@
|
|||
{"id": "Qwen/Qwen3-Embedding-8B", "name": "Qwen3 Embedding 8B", "type": "Embedding", "dimension": 4096, "source": "https://api-docs.siliconflow.cn/docs/api/embeddings-post"},
|
||||
{"id": "Qwen/Qwen3-Embedding-4B", "name": "Qwen3 Embedding 4B", "type": "Embedding", "dimension": 2560, "source": "https://api-docs.siliconflow.cn/docs/api/embeddings-post"},
|
||||
{"id": "Qwen/Qwen3-Embedding-0.6B", "name": "Qwen3 Embedding 0.6B", "type": "Embedding", "dimension": 1024, "source": "https://api-docs.siliconflow.cn/docs/api/embeddings-post"},
|
||||
{"id": "BAAI/bge-reranker-v2-m3", "name": "BGE Reranker v2 M3", "type": "Rerank", "source": "https://docs.siliconflow.cn/cn/api-reference/rerank/create-rerank"},
|
||||
{"id": "Qwen/Qwen3-Reranker-8B", "name": "Qwen3 Reranker 8B", "type": "Rerank", "source": "https://docs.siliconflow.cn/cn/api-reference/rerank/create-rerank"},
|
||||
{"id": "BAAI/bge-reranker-v2-m3", "name": "BGE Reranker v2 M3", "type": "Rerank", "source": "https://api-docs.siliconflow.cn/docs/api/rerank-post"},
|
||||
{"id": "Qwen/Qwen3-Reranker-8B", "name": "Qwen3 Reranker 8B", "type": "Rerank", "source": "https://api-docs.siliconflow.cn/docs/api/rerank-post"},
|
||||
{"id": "FunAudioLLM/SenseVoiceSmall", "name": "SenseVoice Small", "type": "ASR", "source": "https://api-docs.siliconflow.cn/docs/api/audio-transcriptions-post"},
|
||||
{"id": "TeleAI/TeleSpeechASR", "name": "TeleSpeechASR", "type": "ASR", "source": "https://api-docs.siliconflow.cn/docs/api/audio-transcriptions-post"}
|
||||
]
|
||||
|
|
|
|||
|
|
@ -54,7 +54,9 @@
|
|||
// (the API key field carries the access key; the secret key, region and
|
||||
// instruction come from the rerank-only extra fields). Documented models
|
||||
// are doubao-seed-rerank and base-multilingual-rerank
|
||||
// (https://www.volcengine.com/docs/84313/1254474);
|
||||
// (https://www.volcengine.com/docs/84313/1254474). datas takes at most
|
||||
// 200 items, and the default instruction is the console's, verbatim, as
|
||||
// the page asks for results that match the console;
|
||||
// - Ark additionally serves an Anthropic Messages surface at
|
||||
// https://ark.cn-beijing.volces.com/api/compatible/v1 with x-api-key
|
||||
// auth. This package configures OpenAI Chat Completions, which is the
|
||||
|
|
@ -68,10 +70,6 @@
|
|||
// saying whether k is 1000 or 1024, so the context windows in models.json are
|
||||
// left at their present values.
|
||||
//
|
||||
// unverified: the rerank instruction default here capitalises Document /
|
||||
// Query while the console default in the docs is lower case; it is kept in
|
||||
// sync with internal/models/rerank instead.
|
||||
//
|
||||
// unverified: prices in models.json come from the Ark pricing console, which
|
||||
// the public docs do not render.
|
||||
package volcengine
|
||||
|
|
@ -179,7 +177,7 @@ func init() {
|
|||
Label: "Rerank Instruction",
|
||||
Labels: map[string]string{"zh-CN": "重排指令"},
|
||||
Type: "string",
|
||||
Default: "Whether the Document answers the Query or matches the content retrieval intent",
|
||||
Default: "Whether the document answers the query or matches the content retrieval intent",
|
||||
Placeholder: "Instruction passed to the rerank model",
|
||||
Placeholders: map[string]string{
|
||||
"zh-CN": "传给重排模型的 instruction",
|
||||
|
|
@ -223,7 +221,11 @@ func init() {
|
|||
MaxBatchSize: catalog.Ptr(1),
|
||||
},
|
||||
Rerank: catalog.RerankCompat{
|
||||
MaxDocuments: catalog.Ptr(50),
|
||||
// datas "数组长度不超过 200"
|
||||
// (https://docs.volcengine.com/docs/vector_database_vikingdb/Rerank).
|
||||
// The pre-catalog client split at 50, a constant of its own
|
||||
// rather than a documented ceiling.
|
||||
MaxDocuments: catalog.Ptr(200),
|
||||
MaxConcurrency: catalog.Ptr(4),
|
||||
},
|
||||
OpenAICompletions: catalog.OpenAICompletionsCompat{
|
||||
|
|
|
|||
|
|
@ -269,6 +269,12 @@ make model-catalog-diff VENDOR=deepseek # 只看一家
|
|||
3. **`api.openai.com` 的一方流量改走 Responses 协议**(`PreferAPI` 只对官方域生效)。各类中转 / 网关仍走 Chat Completions,`parity` 包里有断言钉住这一点。
|
||||
4. **7 家厂商的输出上限字段按文档纠正**:hunyuan、modelscope、qiniu、requesty、longcat、novita 由 `max_completion_tokens` 改回 `max_tokens`,moonshot 反向改为 `max_completion_tokens`。每一处在 `internal/models/parity/parity_test.go` 里都记了变更理由与厂商文档。aliyun 保持 `max_completion_tokens` 不变:兼容模式两个字段都收,但 DashScope 的参数表已经把 `max_tokens` 标为即将废弃并指名了继任者。
|
||||
|
||||
Rerank 行(逐厂商的出站请求由 `internal/models/rerank/wire_test.go` 钉住,含火山与 LKEAP 两个签名 SDK 客户端):
|
||||
|
||||
1. **OpenAI 不再出现在 rerank 的厂商列表里**。OpenAI 的 API 没有 rerank 接口,在这里建的行只会 404;架在 OpenAI 风格地址后面、自带 rerank 的中转请建成 generic 行。已有的行照常解析。
|
||||
2. **火山 rerank 每次最多 200 条**(文档:datas「数组长度不超过 200」),旧实现按自己定的 50 条切分。
|
||||
3. **火山 rerank 的默认指令改为控制台原文** `Whether the document answers the query or matches the content retrieval intent`,文档要求「如需对齐控制台效果,请使用相同指令」;旧默认值把 Document / Query 写成了大写。已在额外字段里保存了指令的行不受影响。
|
||||
|
||||
Embedding 行也有几处按厂商文档纠正的行为变化(逐厂商的出站请求由 `internal/models/embedding/wire_test.go` 钉住):
|
||||
|
||||
1. **托管厂商不再收到 `truncate_prompt_tokens`**。此前所有 OpenAI 兼容厂商都被塞了一个 511,它只是 vLLM 的扩展参数;`generic`、`gpustack` 照旧发送。
|
||||
|
|
|
|||
Loading…
Reference in New Issue