| 1 | package openai |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "errors" |
| 7 | "fmt" |
| 8 | "io" |
| 9 | "net/http" |
| 10 | "slices" |
| 11 | "sort" |
| 12 | "strings" |
| 13 | "time" |
| 14 | |
| 15 | "reasonix/internal/netclient" |
| 16 | "reasonix/internal/provider" |
| 17 | ) |
| 18 | |
| 19 | type modelFetchStatusError struct { |
| 20 | status int |
| 21 | body string |
| 22 | } |
| 23 | |
| 24 | type ModelFetchAuthMode string |
| 25 | |
| 26 | const ( |
| 27 | ModelFetchAuthAuto ModelFetchAuthMode = "" |
| 28 | ModelFetchAuthBearer ModelFetchAuthMode = "bearer" |
| 29 | ModelFetchAuthXAPIKey ModelFetchAuthMode = "x-api-key" |
| 30 | |
| 31 | // fetchModelsMaxBody caps the response body read from a model-list |
| 32 | // endpoint. Large providers like OpenRouter return ~530 KB for 338 |
| 33 | // models; 2 MiB leaves headroom while keeping memory bounded. |
| 34 | fetchModelsMaxBody = 2 << 20 // 2 MiB |
| 35 | ) |
| 36 | |
| 37 | type FetchModelsOptions struct { |
| 38 | Headers map[string]string |
| 39 | AuthMode ModelFetchAuthMode |
| 40 | // Proxy routes the model-list request through the same transport policy as |
| 41 | // chat requests, so a broken proxy surfaces at setup time instead of only |
| 42 | // stalling the first chat turn later (#9560). |
| 43 | Proxy netclient.ProxySpec |
| 44 | } |
| 45 | |
| 46 | func (e modelFetchStatusError) Error() string { |
| 47 | return fmt.Sprintf("fetch models: status %d: %s", e.status, strings.TrimSpace(e.body)) |
| 48 | } |
| 49 | |
| 50 | // IsModelFetchEndpointMiss reports whether a model-list request reached a |
| 51 | // plausible endpoint path that the provider does not implement. |
| 52 | func IsModelFetchEndpointMiss(err error) bool { |
| 53 | var statusErr modelFetchStatusError |
| 54 | if !errors.As(err, &statusErr) { |
| 55 | return false |
| 56 | } |
| 57 | return statusErr.status == http.StatusNotFound || statusErr.status == http.StatusMethodNotAllowed |
| 58 | } |
| 59 | |
| 60 | // FetchModels calls the OpenAI-compatible GET /models endpoint and returns the |
| 61 | // available model IDs. |
| 62 | func FetchModels(ctx context.Context, baseURL, apiKey string, headers map[string]string) ([]string, error) { |
| 63 | return FetchModelsWithOptions(ctx, baseURL, apiKey, FetchModelsOptions{Headers: headers}) |
| 64 | } |
| 65 | |
| 66 | // FetchModelCatalog calls the OpenAI-compatible model endpoint and returns |
| 67 | // model-level capability metadata. The adapter deliberately uses a |
| 68 | // unknown capability when an endpoint omits capability fields; |
| 69 | // callers must never infer image support from a model name. |
| 70 | func FetchModelCatalog(ctx context.Context, baseURL, apiKey string, headers map[string]string) ([]provider.ModelInfo, error) { |
| 71 | return FetchModelCatalogWithOptions(ctx, baseURL, apiKey, FetchModelsOptions{Headers: headers}) |
| 72 | } |
| 73 | |
| 74 | // FetchModelsWithOptions calls the OpenAI-compatible GET /models endpoint and |
| 75 | // returns the available model IDs. |
| 76 | func FetchModelsWithOptions(ctx context.Context, baseURL, apiKey string, opts FetchModelsOptions) ([]string, error) { |
| 77 | catalog, err := FetchModelCatalogWithOptions(ctx, baseURL, apiKey, opts) |
| 78 | if err != nil { |
| 79 | return nil, err |
| 80 | } |
| 81 | ids := make([]string, 0, len(catalog)) |
| 82 | for _, model := range catalog { |
| 83 | ids = append(ids, model.ID) |
| 84 | } |
| 85 | return ids, nil |
| 86 | } |
| 87 | |
| 88 | // FetchModelCatalogWithOptions is the metadata-preserving form of |
| 89 | // FetchModelsWithOptions. It keeps the existing request/auth/size behavior. |
| 90 | func FetchModelCatalogWithOptions(ctx context.Context, baseURL, apiKey string, opts FetchModelsOptions) ([]provider.ModelInfo, error) { |
| 91 | transport, err := netclient.NewTransport(opts.Proxy, netclient.TransportOptions{}) |
| 92 | if err != nil { |
| 93 | return nil, fmt.Errorf("fetch models: network: %w", err) |
| 94 | } |
| 95 | cli := &http.Client{Timeout: 10 * time.Second, Transport: transport} |
| 96 | url := strings.TrimRight(baseURL, "/") |
| 97 | if !strings.HasSuffix(url, "/models") { |
| 98 | url += "/models" |
| 99 | } |
| 100 | |
| 101 | req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil) |
| 102 | if err != nil { |
| 103 | return nil, fmt.Errorf("fetch models: build request: %w", err) |
| 104 | } |
| 105 | applyModelFetchAPIKeyHeader(req.Header, baseURL, apiKey, opts.AuthMode) |
| 106 | req.Header.Set("Accept", "application/json") |
| 107 | applyCustomHeaders(req.Header, opts.Headers) |
| 108 | |
| 109 | resp, err := cli.Do(req) |
| 110 | if err != nil { |
| 111 | return nil, fmt.Errorf("fetch models: request failed: %w", err) |
| 112 | } |
| 113 | defer resp.Body.Close() |
| 114 | |
| 115 | body, err := io.ReadAll(io.LimitReader(resp.Body, fetchModelsMaxBody+1)) |
| 116 | if err != nil { |
| 117 | return nil, fmt.Errorf("fetch models: read response: %w", err) |
| 118 | } |
| 119 | if len(body) > fetchModelsMaxBody { |
| 120 | return nil, fmt.Errorf("fetch models: response too large (exceeds %d bytes)", fetchModelsMaxBody) |
| 121 | } |
| 122 | |
| 123 | if resp.StatusCode != http.StatusOK { |
| 124 | return nil, modelFetchStatusError{status: resp.StatusCode, body: truncateFetchBody(string(body))} |
| 125 | } |
| 126 | |
| 127 | var result struct { |
| 128 | Data []json.RawMessage `json:"data"` |
| 129 | } |
| 130 | if err := json.Unmarshal(body, &result); err != nil { |
| 131 | return nil, fmt.Errorf("fetch models: decode response: %w", err) |
| 132 | } |
| 133 | |
| 134 | modelsByID := make(map[string]provider.ModelInfo, len(result.Data)) |
| 135 | conflicts := make(map[string]bool) |
| 136 | for _, raw := range result.Data { |
| 137 | model := parseModelInfo(baseURL, raw) |
| 138 | if model.ID == "" { |
| 139 | continue |
| 140 | } |
| 141 | if previous, exists := modelsByID[model.ID]; exists { |
| 142 | switch { |
| 143 | case conflicts[model.ID]: |
| 144 | model.InputModalities = nil |
| 145 | case model.InputModalities == nil: |
| 146 | model.InputModalities = previous.InputModalities |
| 147 | case previous.InputModalities != nil && !sameModalities(previous.InputModalities, model.InputModalities): |
| 148 | // A conflict stays unknown for the rest of this response, regardless |
| 149 | // of duplicate ordering. Missing metadata is not a negative fact. |
| 150 | conflicts[model.ID] = true |
| 151 | model.InputModalities = nil |
| 152 | } |
| 153 | } |
| 154 | modelsByID[model.ID] = model |
| 155 | } |
| 156 | models := make([]provider.ModelInfo, 0, len(modelsByID)) |
| 157 | for _, model := range modelsByID { |
| 158 | models = append(models, model) |
| 159 | } |
| 160 | sort.Slice(models, func(i, j int) bool { return models[i].ID < models[j].ID }) |
| 161 | return models, nil |
| 162 | } |
| 163 | |
| 164 | func sameModalities(a, b []provider.ModelModality) bool { |
| 165 | return len(a) == len(b) && !slices.ContainsFunc(a, func(m provider.ModelModality) bool { return !slices.Contains(b, m) }) |
| 166 | } |
| 167 | |
| 168 | func parseModelInfo(baseURL string, raw json.RawMessage) provider.ModelInfo { |
| 169 | var entry map[string]json.RawMessage |
| 170 | if json.Unmarshal(raw, &entry) != nil { |
| 171 | return provider.ModelInfo{} |
| 172 | } |
| 173 | var rawID string |
| 174 | _ = json.Unmarshal(entry["id"], &rawID) |
| 175 | id := normalizeModelID(baseURL, rawID) |
| 176 | if id == "" { |
| 177 | return provider.ModelInfo{} |
| 178 | } |
| 179 | modalities, _ := parseModalities(entry) |
| 180 | return provider.ModelInfo{ID: id, InputModalities: modalities} |
| 181 | } |
| 182 | |
| 183 | func parseModalities(entry map[string]json.RawMessage) ([]provider.ModelModality, bool) { |
| 184 | // Canonical and nested array fields are ordered before compatibility |
| 185 | // aliases. Presence with an invalid value is treated as an unsafe |
| 186 | // declaration and therefore stays unknown. |
| 187 | for _, key := range []string{"input_modalities"} { |
| 188 | if raw, present := entry[key]; present { |
| 189 | return decodeModalities(raw) |
| 190 | } |
| 191 | } |
| 192 | if raw, present := entry["modalities"]; present { |
| 193 | var nested map[string]json.RawMessage |
| 194 | if json.Unmarshal(raw, &nested) == nil { |
| 195 | if input, ok := nested["input"]; ok { |
| 196 | return decodeModalities(input) |
| 197 | } |
| 198 | } |
| 199 | return nil, false |
| 200 | } |
| 201 | if raw, present := entry["capabilities"]; present { |
| 202 | var nested map[string]json.RawMessage |
| 203 | if json.Unmarshal(raw, &nested) == nil { |
| 204 | if input, ok := nested["input_modalities"]; ok { |
| 205 | return decodeModalities(input) |
| 206 | } |
| 207 | if vision, ok := nested["vision"]; ok { |
| 208 | return decodeVisionBool(vision) |
| 209 | } |
| 210 | } |
| 211 | return nil, false |
| 212 | } |
| 213 | for _, key := range []string{"supports_vision", "vision"} { |
| 214 | if raw, present := entry[key]; present { |
| 215 | return decodeVisionBool(raw) |
| 216 | } |
| 217 | } |
| 218 | return nil, false |
| 219 | } |
| 220 | |
| 221 | func decodeModalities(raw json.RawMessage) ([]provider.ModelModality, bool) { |
| 222 | var values []string |
| 223 | if json.Unmarshal(raw, &values) != nil || len(values) == 0 { |
| 224 | return nil, false |
| 225 | } |
| 226 | seen := map[provider.ModelModality]bool{} |
| 227 | out := make([]provider.ModelModality, 0, len(values)) |
| 228 | for _, value := range values { |
| 229 | modality := provider.ModelModality(strings.ToLower(strings.TrimSpace(value))) |
| 230 | if modality != provider.ModalityText && modality != provider.ModalityImage { |
| 231 | return nil, false |
| 232 | } |
| 233 | if !seen[modality] { |
| 234 | seen[modality] = true |
| 235 | out = append(out, modality) |
| 236 | } |
| 237 | } |
| 238 | // Stable order also makes duplicate merging independent of array order. |
| 239 | if len(out) == 2 && out[0] == provider.ModalityImage { |
| 240 | out[0], out[1] = out[1], out[0] |
| 241 | } |
| 242 | return out, len(out) > 0 |
| 243 | } |
| 244 | |
| 245 | func decodeVisionBool(raw json.RawMessage) ([]provider.ModelModality, bool) { |
| 246 | var vision *bool |
| 247 | if json.Unmarshal(raw, &vision) != nil || vision == nil { |
| 248 | return nil, false |
| 249 | } |
| 250 | if *vision { |
| 251 | return []provider.ModelModality{provider.ModalityText, provider.ModalityImage}, true |
| 252 | } |
| 253 | return []provider.ModelModality{provider.ModalityText}, true |
| 254 | } |
| 255 | |
| 256 | func applyModelFetchAPIKeyHeader(h http.Header, baseURL, apiKey string, mode ModelFetchAuthMode) { |
| 257 | apiKey = strings.TrimSpace(apiKey) |
| 258 | if apiKey == "" { |
| 259 | return |
| 260 | } |
| 261 | switch mode { |
| 262 | case ModelFetchAuthBearer: |
| 263 | h.Set("Authorization", "Bearer "+apiKey) |
| 264 | case ModelFetchAuthXAPIKey: |
| 265 | h.Set("x-api-key", apiKey) |
| 266 | default: |
| 267 | applyAPIKeyHeader(h, baseURL, apiKey) |
| 268 | } |
| 269 | } |
| 270 | |
| 271 | func truncateFetchBody(body string) string { |
| 272 | body = strings.TrimSpace(body) |
| 273 | const max = 512 |
| 274 | if len([]rune(body)) <= max { |
| 275 | return body |
| 276 | } |
| 277 | r := []rune(body) |
| 278 | return string(r[:max]) + "..." |
| 279 | } |
| 280 |