Download internal/server/embeddings.go from a3216/gcli2api: direct link, hf CLI and curl.
- Browser
- Download file 9.25 kB
-
https://huggingface.co/spaces/a3216/gcli2api/resolve/main/internal/server/embeddings.go
- Command line
-
hf download hf://spaces/a3216/gcli2api/internal/server/embeddings.go
-
curl -L -o embeddings.go https://huggingface.co/spaces/a3216/gcli2api/resolve/main/internal/server/embeddings.go
9.25 kB
| // embeddings.go POST /v1/embeddings:OpenAI 兼容嵌入端点(实验特性,默认关闭)。 | |
| // | |
| // 上游嵌入与对话是两套注册表(/v3/config 不含嵌入模型),实测唯一可用的是 bge-m3 | |
| // (1024 维)。因此本端点: | |
| // - 默认关闭:关着时回 403 embeddings_disabled(不是 404——功能存在,只是没开)。 | |
| // 开关是 features.embeddings_enabled,面板保存后热生效(livecfg 快照)。 | |
| // - 模型名原样透传给上游,不替客户端改模型;未知模型由上游回 11351,本端映射成 | |
| // OpenAI 语义的 404 model_not_found,并在文案里点出已知可用模型。 | |
| // - 回程把上游 data[] 补齐 OpenAI 契约(object/index);条数与输入不符时宁可不返回, | |
| // 也不能错位配对(向量错位比报错更难查)。 | |
| // - 取号与失败处置与 /v1/search、/v1/images 同一套:单次取号、换号上限 3、 | |
| // 请求级错误(参数/内容/模型)fail-fast 不烧全池。 | |
| package server | |
| import ( | |
| "encoding/json" | |
| "errors" | |
| "fmt" | |
| "io" | |
| "log" | |
| "math" | |
| "net/http" | |
| "strconv" | |
| "strings" | |
| "github.com/misakano7545/workbuddy2api-panel/internal/logfmt" | |
| "github.com/misakano7545/workbuddy2api-panel/internal/proxypool" | |
| "github.com/misakano7545/workbuddy2api-panel/internal/upstream" | |
| ) | |
| // embeddingsAttempts 失败换号上限(与 search/images 同量级)。 | |
| const embeddingsAttempts = 3 | |
| type embeddingsRequest struct { | |
| Model string `json:"model"` | |
| // Input 支持 OpenAI 的两种常用形态:单条字符串 / 字符串数组。 | |
| // 分词数组([][]int)不接——上游只认文本,转发无意义。 | |
| Input any `json:"input"` | |
| // Realm 可选,直接指定取号域(优先于 model 前缀)。 | |
| Realm string `json:"realm"` | |
| } | |
| // normalizeEmbeddingInput 把 input 归一成字符串数组。 | |
| // 空串与空数组一律报错:上游对空输入没有有意义的行为,交给它只会换来一条难懂的 500。 | |
| func normalizeEmbeddingInput(v any) ([]string, error) { | |
| switch t := v.(type) { | |
| case nil: | |
| return nil, errors.New("缺少 input 字段") | |
| case string: | |
| if t == "" { | |
| return nil, errors.New("input 不能为空") | |
| } | |
| return []string{t}, nil | |
| case []any: | |
| if len(t) == 0 { | |
| return nil, errors.New("input 不能为空数组") | |
| } | |
| out := make([]string, 0, len(t)) | |
| for i, item := range t { | |
| s, ok := item.(string) | |
| if !ok { | |
| return nil, fmt.Errorf("input[%d] 必须是字符串", i) | |
| } | |
| if s == "" { | |
| return nil, fmt.Errorf("input[%d] 不能为空串", i) | |
| } | |
| out = append(out, s) | |
| } | |
| return out, nil | |
| default: | |
| return nil, errors.New("input 只支持字符串或字符串数组") | |
| } | |
| } | |
| // normalizeEmbeddingsResponse 把上游原文补成 OpenAI 契约: | |
| // data[i] 必须有 object="embedding" 与下标记;顶层 object="list"。 | |
| // 上游缺 index 时按数组顺序补(上游返回顺序即输入顺序,这是 OpenAI 的约定)。 | |
| // 条数与输入数不符时返回错误——错位配对比报错更难查。 | |
| func normalizeEmbeddingsResponse(raw []byte, want int) ([]byte, error) { | |
| var obj map[string]json.RawMessage | |
| if err := json.Unmarshal(raw, &obj); err != nil { | |
| return nil, fmt.Errorf("上游响应不是 JSON: %w", err) | |
| } | |
| var items []map[string]any | |
| if err := json.Unmarshal(obj["data"], &items); err != nil { | |
| return nil, errors.New("上游响应缺少 data 向量数组") | |
| } | |
| if len(items) != want { | |
| return nil, fmt.Errorf("上游返回 %d 条向量,与输入 %d 条不符", len(items), want) | |
| } | |
| out := make([]map[string]any, len(items)) | |
| seen := make(map[int]bool, len(items)) | |
| for i, it := range items { | |
| vector, ok := it["embedding"].([]any) | |
| if !ok || len(vector) == 0 { | |
| return nil, fmt.Errorf("上游 data[%d] 缺少非空数值向量", i) | |
| } | |
| for _, value := range vector { | |
| if _, ok := value.(float64); !ok { | |
| return nil, fmt.Errorf("上游 data[%d] 向量含非数值元素", i) | |
| } | |
| } | |
| index := i | |
| if value, present := it["index"]; present { | |
| n, ok := value.(float64) | |
| if !ok || n < 0 || n >= float64(want) || math.Trunc(n) != n { | |
| return nil, fmt.Errorf("上游 data[%d] 下标无效", i) | |
| } | |
| index = int(n) | |
| } | |
| if seen[index] { | |
| return nil, fmt.Errorf("上游向量下标 %d 重复", index) | |
| } | |
| seen[index] = true | |
| it["object"] = "embedding" | |
| it["index"] = index | |
| out[i] = it | |
| } | |
| data, err := json.Marshal(out) | |
| if err != nil { | |
| return nil, err | |
| } | |
| obj["data"] = data | |
| obj["object"] = json.RawMessage(`"list"`) | |
| // 其余字段(model / usage / 未来新增)原样保留。 | |
| return json.Marshal(obj) | |
| } | |
| func (h *Handler) embeddings(w http.ResponseWriter, r *http.Request) { | |
| if !h.loadLive().EmbeddingsEnabled { | |
| writeOpenAIError(w, http.StatusForbidden, "embeddings_disabled", | |
| "嵌入端点未开启:请在面板配置里打开 features.embeddings_enabled(默认关闭)") | |
| return | |
| } | |
| raw, err := io.ReadAll(io.LimitReader(r.Body, (8<<20)+1)) | |
| if err != nil { | |
| writeOpenAIError(w, http.StatusBadRequest, "invalid_request", "read body: "+err.Error()) | |
| return | |
| } | |
| if len(raw) > 8<<20 { | |
| writeOpenAIError(w, http.StatusRequestEntityTooLarge, "invalid_request", "request body exceeds 8 MiB") | |
| return | |
| } | |
| var req embeddingsRequest | |
| if err := json.Unmarshal(raw, &req); err != nil { | |
| writeOpenAIError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error()) | |
| return | |
| } | |
| model := strings.TrimSpace(req.Model) | |
| if model == "" { | |
| writeOpenAIError(w, http.StatusBadRequest, "invalid_request", "model is required") | |
| return | |
| } | |
| inputs, err := normalizeEmbeddingInput(req.Input) | |
| if err != nil { | |
| writeOpenAIError(w, http.StatusBadRequest, "invalid_request", err.Error()) | |
| return | |
| } | |
| // 域:realm 显式优先,其次模型自带前缀(cn: / global:),都没有 → cn。 | |
| // 与对话不同,**剥掉**前缀再发给上游:上游目录里的模型 id 不带域前缀(bge-m3)。 | |
| realm, bare := resolveModel(model) | |
| if req.Realm != "" { | |
| realm = req.Realm | |
| } | |
| if realm != "global" { | |
| realm = "cn" | |
| } | |
| if bare == "" { | |
| bare = model | |
| } | |
| if !h.allowedModel(realm, bare) || h.hiddenModel(realm, bare) { | |
| writeOpenAIError(w, http.StatusBadRequest, "model_unavailable", "model is unavailable by gateway configuration, try another model") | |
| return | |
| } | |
| tried := map[string]bool{} | |
| var lastErr error | |
| for attempt := 0; attempt < embeddingsAttempts; attempt++ { | |
| a := h.cfg.Pool.PickExcludingForRealm(tried, "", realm) | |
| if a == nil { | |
| break | |
| } | |
| tried[a.UID] = true | |
| data, err := h.cfg.Upstream.Embeddings(r.Context(), a, bare, inputs) | |
| if err == nil { | |
| out, nerr := normalizeEmbeddingsResponse(data, len(inputs)) | |
| if nerr != nil { | |
| // 形状不对是上游的问题,但换号大概率一样:直接 502,不烧池。 | |
| log.Printf("embeddings %s: 响应形状异常: %v", logfmt.Label(a.UID, a.Nickname), nerr) | |
| writeOpenAIError(w, http.StatusBadGateway, "upstream_error", nerr.Error()) | |
| return | |
| } | |
| w.Header().Set("Content-Type", "application/json") | |
| _, _ = w.Write(out) | |
| return | |
| } | |
| lastErr = err | |
| if proxypool.IsError(err) { | |
| break | |
| } | |
| // 模型不存在:换号无意义(注册表是按模型不是按账号),且要给出 OpenAI 的 404 语义。 | |
| if errors.Is(err, upstream.ErrEmbeddingModelNotFound) { | |
| writeOpenAIError(w, http.StatusNotFound, "model_not_found", | |
| fmt.Sprintf("嵌入模型 %q 不存在,当前上游嵌入模型仅支持 %q", bare, upstream.EmbeddingModelBGE)) | |
| return | |
| } | |
| var ue *upstream.Error | |
| if errors.As(err, &ue) { | |
| // 6020/6021 渠道级队列:上游对本形态本身回 429(6020 带 data.retry_after)。 | |
| // 与 chat 链路同口径(PR #108):429 + Retry-After,不换号(全局队列,换谁都撞), | |
| // 也不罚号(applyErrorPolicy 的 ErrUpstreamBusy 分支零动作)。 | |
| if ue.Kind == upstream.ErrUpstreamBusy { | |
| if ue.RetryAfter > 0 { | |
| w.Header().Set("Retry-After", strconv.Itoa(int(math.Ceil(ue.RetryAfter.Seconds())))) | |
| } | |
| writeOpenAIError(w, http.StatusTooManyRequests, "upstream_busy", ue.Msg) | |
| return | |
| } | |
| h.applyErrorPolicy(a.UID, ue.Kind, ue.Msg, "", ue) | |
| // 请求级错误(参数/内容/模型权限):换号必然同样被拒,fail-fast 保住池子。 | |
| if requestLevelFailFastKind(ue.Kind) { | |
| status := ue.Status | |
| if status < 400 { | |
| status = http.StatusBadRequest | |
| } | |
| log.Printf("embeddings realm=%s: 请求级错误 kind=%v,不再换号", realm, ue.Kind) | |
| writeOpenAIError(w, status, "upstream_error", ue.Msg) | |
| return | |
| } | |
| } | |
| log.Printf("embeddings %s realm=%s attempt=%d: %v", logfmt.Label(a.UID, a.Nickname), realm, attempt+1, err) | |
| } | |
| status, code := http.StatusBadGateway, "upstream_error" | |
| var ue *upstream.Error | |
| switch { | |
| case proxypool.IsError(lastErr): | |
| status, code = http.StatusServiceUnavailable, "proxy_unavailable" | |
| case len(tried) == 0: | |
| status, code = http.StatusServiceUnavailable, "no_account" | |
| lastErr = fmt.Errorf("没有可用的 %s 账号", realm) | |
| case errors.As(lastErr, &ue) && ue.Status >= 400: | |
| status = ue.Status | |
| } | |
| log.Printf("embeddings realm=%s: 全部尝试失败(%d 个账号):%v", realm, len(tried), lastErr) | |
| writeOpenAIError(w, status, code, lastErr.Error()) | |
| } | |