plugin-ai-api 1.0.2 → 1.0.4

package/dist/client/plugin.d.ts

@@ -1,6 +0,0 @@
-import { Plugin } from '@nocobase/client';
-export declare class PluginAiApiClient extends Plugin {
-    load(): Promise<void>;
-    t: (key: string) => string;
-}
-export default PluginAiApiClient;

package/dist/index.d.ts

@@ -1,2 +0,0 @@
-export * from './server';
-export { default } from './server';

package/dist/server/collections/ai-api-config.d.ts

@@ -1,2 +0,0 @@
-declare const _default: any;
-export default _default;

package/dist/server/collections/ai-api-role-permissions.d.ts

@@ -1,2 +0,0 @@
-declare const _default: any;
-export default _default;

package/dist/server/index.d.ts

@@ -1 +0,0 @@
-export { default } from './plugin';

package/dist/server/middleware/rate-limit.d.ts

@@ -1,18 +0,0 @@
-import { RateLimiter } from '../utils/rate-limiter';
-/**
- * Creates a rate-limiting check function for use in the AI API router.
- *
- * Must be called AFTER authenticateBearer() so ctx.state.currentUser is set.
- * Reads rateLimitPerMinute from aiApiConfig on each request (cheap single-row read,
- * allows config changes to take effect immediately without restart).
- * Falls back to 60 req/min if the config record is missing or the field is 0/null.
- *
- * Returns false (and writes the 429 response) when the rate limit is exceeded.
- * Returns true when the request is allowed.
- *
- * Sets OpenAI-compatible rate limit response headers on every request:
- *   X-RateLimit-Limit: <limit>
- *   X-RateLimit-Remaining: <remaining> (on 429: 0)
- *   Retry-After: <seconds>  (on 429 only)
- */
-export declare function createRateLimitMiddleware(limiter: RateLimiter): (ctx: Context) => Promise<boolean>;

package/dist/server/middleware/role-permission.d.ts

@@ -1,21 +0,0 @@
-import { Context } from '@nocobase/actions';
-/**
- * Check whether the authenticated role is allowed to use the AI API.
- * Loads the permission record and stores it in ctx.state.aiApiRolePermission.
- *
- * Returns true if access is allowed (caller may proceed).
- * Returns false if access is denied (403 already written to ctx, caller must return).
- *
- * The 'root' and 'admin' roles always bypass the check.
- */
-export declare function checkRolePermission(ctx: Context): Promise<boolean>;
-/**
- * Check whether the current role is allowed to use a specific AI Employee.
- * Must be called after checkRolePermission (so ctx.state.aiApiRolePermission is set).
- *
- * Returns true when:
- * - Role is admin/root (no permission record stored)
- * - allowAllEmployees is true
- * - The employeeUsername is in the allowedEmployees list
- */
-export declare function checkEmployeeAccess(ctx: Context, employeeUsername: string): boolean;

package/dist/server/plugin.d.ts

@@ -1,18 +0,0 @@
-import { Plugin } from '@nocobase/server';
-import { RateLimiter } from './utils/rate-limiter';
-export declare class PluginAiApiServer extends Plugin {
-    /**
-     * Singleton rate limiter — lives for the entire plugin lifetime, shared across all requests.
-     * Uses a 1-minute sliding window to enforce rateLimitPerMinute from aiApiConfig.
-     */
-    rateLimiter: RateLimiter;
-    private gcInterval;
-    afterAdd(): Promise<void>;
-    beforeLoad(): Promise<void>;
-    load(): Promise<void>;
-    install(): Promise<void>;
-    afterEnable(): Promise<void>;
-    afterDisable(): Promise<void>;
-    remove(): Promise<void>;
-}
-export default PluginAiApiServer;

package/dist/server/resource/ai-api-config.d.ts

@@ -1,10 +0,0 @@
-import { ResourceOptions } from '@nocobase/resourcer';
-/**
- * Resource for managing AI API configuration via NocoBase admin UI.
- * Singleton config pattern (same as aiSettings in plugin-ai).
- *
- * Uses custom action name 'save' instead of 'update' because NocoBase's
- * built-in middleware requires filter/filterByTk for the standard 'update' action.
- */
-declare const aiApiConfigResource: ResourceOptions;
-export default aiApiConfigResource;

package/dist/server/routes/agent-completions.d.ts

@@ -1,34 +0,0 @@
-import { Context } from '@nocobase/actions';
-import type PluginAiApiServer from '../plugin';
-/**
- * POST /api/ai-llm/v1/chat/completions (agent mode)
- *
- * Runs the full AI Employee pipeline by directly instantiating AIEmployee
- * from plugin-ai. This provides TRUE real-time streaming — no buffering.
- *
- * ## Architecture (vs old approach)
- *
- * OLD (fake streaming, HTTP loopback):
- *   Client → ai-api → HTTP to localhost → aiConversations:sendMessages
- *                   ← buffers entire response ←
- *   Client ← 20-char fake chunks ← re-emit
- *
- * NEW (true streaming, direct instantiation):
- *   Client → ai-api → AIEmployee(ctx) ← writes directly to ctx.res
- *   We intercept ctx.res.write to translate NocoBase SSE → OpenAI SSE
- *   Client ← real-time OpenAI SSE chunks ← write() intercept
- *
- * ## NocoBase → OpenAI SSE translation
- *
- * NocoBase emits: `data: {"type":"content","body":"chunk text"}\n\n`
- * We emit:        `data: {"choices":[{"delta":{"content":"chunk text"},...}]}\n\n`
- *
- * Other NocoBase event types (tool_calls, stream_start, etc.) are silently
- * ignored — they are NocoBase-internal events not part of the OpenAI protocol.
- *
- * ## Known limitation
- * Token usage is always { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 }
- * in agent mode. ResponseMetadataCollector is private inside AIEmployee with no
- * public accessor. Modifying plugin-ai is out of scope for this plugin.
- */
-export declare function handleAgentCompletions(ctx: Context, plugin: PluginAiApiServer): Promise<void>;

package/dist/server/routes/auth.d.ts

@@ -1,8 +0,0 @@
-import { Context } from '@nocobase/actions';
-/**
- * Authentication middleware for /api/ai-llm/v1/* routes.
- *
- * Validates Bearer token against NocoBase's apiKeys collection.
- * Sets ctx.state.currentUser for downstream handlers.
- */
-export declare function authenticateBearer(ctx: Context): Promise<boolean>;

package/dist/server/routes/chat-completions.d.ts

@@ -1,9 +0,0 @@
-import { Context } from '@nocobase/actions';
-import type PluginAiApiServer from '../plugin';
-/**
- * POST /api/ai-llm/v1/chat/completions
- *
- * Handles OpenAI-compatible chat completion requests.
- * Supports both streaming (SSE) and non-streaming modes.
- */
-export declare function handleChatCompletions(ctx: Context, plugin: PluginAiApiServer): Promise<void>;

package/dist/server/routes/completions.d.ts

@@ -1,10 +0,0 @@
-import { Context } from '@nocobase/actions';
-import type PluginAiApiServer from '../plugin';
-/**
- * POST /api/ai-llm/v1/completions
- *
- * Handles legacy OpenAI text completions API.
- * Converts prompt to messages and delegates to the chat model.
- * Required by LiteLLM and other tools that test via legacy completions endpoint.
- */
-export declare function handleCompletions(ctx: Context, plugin: PluginAiApiServer): Promise<void>;

package/dist/server/routes/embeddings.d.ts

@@ -1,22 +0,0 @@
-import { Context } from '@nocobase/actions';
-import type PluginAiApiServer from '../plugin';
-/**
- * POST /api/ai-llm/v1/embeddings
- *
- * OpenAI-compatible embeddings endpoint.
- *
- * Supported providers (those with an `embedding` field in LLMProviderMeta):
- *   - openai            → OpenAiEmbeddingProvider
- *   - openai-completions → OpenAiEmbeddingProvider
- *   - dashscope         → DashscopeEmbeddingProvider
- *   - google-genai      → GoogleGenAIEmbeddingProvider
- *   - ollama            → OllamaEmbeddingProvider
- *
- * Not supported: anthropic, deepseek, kimi (no embedding provider registered).
- *
- * Limitations:
- * - encoding_format 'base64' is not supported (always returns float arrays)
- * - Token counts always return 0 (LangChain embeddings API doesn't expose this)
- * - Token array input (integer[]) is not supported, only string input
- */
-export declare function handleEmbeddings(ctx: Context, plugin: PluginAiApiServer): Promise<void>;

package/dist/server/routes/models.d.ts

@@ -1,21 +0,0 @@
-import { Context } from '@nocobase/actions';
-import type PluginAiApiServer from '../plugin';
-/**
- * GET /api/ai-llm/v1/models
- *
- * Lists all available models from enabled LLM services.
- * Model IDs use the "serviceName/modelId" format (e.g. "my-openai/gpt-4o")
- * so clients can copy-paste the ID directly into POST /v1/chat/completions
- * without needing to configure a defaultLlmService.
- *
- * Backward compatibility: resolveModelString() in resolve-service.ts still
- * accepts bare model IDs via its 3-tier fallback (defaultLlmService / single service).
- */
-export declare function handleListModels(ctx: Context, plugin: PluginAiApiServer): Promise<void>;
-/**
- * GET /api/ai-llm/v1/models/:model
- *
- * Retrieve a single model by ID.
- * Accepts both "serviceName/modelId" format (new) and bare "modelId" (backward compat).
- */
-export declare function handleGetModel(ctx: Context, modelId: string, plugin: PluginAiApiServer): Promise<void>;

package/dist/server/routes/router.d.ts

@@ -1,25 +0,0 @@
-import type PluginAiApiServer from '../plugin';
-/**
- * Main Koa middleware router for OpenAI-compatible endpoints.
- *
- * Intercepts all requests to /api/ai-llm/v1/* and routes them
- * to the appropriate handler. Runs before NocoBase's resourcer
- * so the URL paths follow OpenAI convention.
- *
- * Features:
- * - CORS support (Access-Control-Allow-Origin: *)
- * - OPTIONS preflight handling (204)
- * - X-Request-Id on every response
- * - Bearer token authentication
- * - Sliding window rate limiting (enforces rateLimitPerMinute from config)
- * - Structured request logging via app.logger
- *
- * Supported endpoints:
- *   POST /v1/chat/completions  — OpenAI chat completions (LLM or agent mode)
- *   POST /v1/completions       — Legacy text completions (LiteLLM compat)
- *   POST /v1/embeddings        — OpenAI embeddings
- *   GET  /v1/models            — List available models
- *   GET  /v1/models/:id        — Get a single model
- *   DELETE /v1/models/:id      — Not implemented (501 stub)
- */
-export declare function createAiLlmRouter(plugin: PluginAiApiServer): (ctx: Context, next: Next) => Promise<any>;

package/dist/server/utils/openai-format.d.ts

@@ -1,99 +0,0 @@
-export interface ParsedModel {
-    llmService: string;
-    modelId: string;
-}
-/**
- * Parse OpenAI-style model string into NocoBase llmService + modelId.
- * Format: "llmServiceName/modelId" (e.g. "my-openai/gpt-4o")
- * If no "/" is present, the entire string is treated as modelId and llmService is empty.
- */
-export declare function parseModelString(model: string): ParsedModel;
-export declare function generateCompletionId(): string;
-export declare function toOpenAIError(statusCode: number, message: string, type?: string, code?: string): {
-    error: {
-        message: string;
-        type: string;
-        param: any;
-        code: string;
-    };
-};
-export declare function toOpenAIResponse(options: {
-    id: string;
-    model: string;
-    content: string;
-    finishReason?: string;
-    usage?: {
-        prompt_tokens?: number;
-        completion_tokens?: number;
-        total_tokens?: number;
-    };
-}): {
-    id: string;
-    object: string;
-    created: number;
-    model: string;
-    system_fingerprint: any;
-    choices: {
-        index: number;
-        message: {
-            role: string;
-            content: string;
-        };
-        logprobs: any;
-        finish_reason: string;
-    }[];
-    usage: {
-        prompt_tokens?: number;
-        completion_tokens?: number;
-        total_tokens?: number;
-    };
-};
-export declare function toOpenAIStreamChunk(options: {
-    id: string;
-    model: string;
-    delta: {
-        role?: string;
-        content?: string;
-    };
-    finishReason?: string | null;
-}): {
-    id: string;
-    object: string;
-    created: number;
-    model: string;
-    system_fingerprint: any;
-    choices: {
-        index: number;
-        delta: {
-            role?: string;
-            content?: string;
-        };
-        logprobs: any;
-        finish_reason: string;
-    }[];
-};
-export declare function toOpenAIEmbeddingsResponse(options: {
-    model: string;
-    embeddings: number[][];
-    promptTokens?: number;
-}): {
-    object: "list";
-    data: {
-        object: "embedding";
-        embedding: number[];
-        index: number;
-    }[];
-    model: string;
-    usage: {
-        prompt_tokens: number;
-        total_tokens: number;
-    };
-};
-/**
- * Format a streaming chunk as an SSE data line.
- */
-export declare function formatSSE(data: any): string;
-/**
- * Format the terminal SSE [DONE] signal.
- */
-export declare function formatSSEDone(): string;

package/dist/server/utils/rate-limiter.d.ts

@@ -1,35 +0,0 @@
-/**
- * In-memory sliding window rate limiter.
- *
- * Stores per-user request timestamps. On each check: prunes timestamps
- * older than the window, counts the remainder, and accepts/rejects.
- *
- * Single-process safe (Node.js event loop). Not distributed.
- * For multi-process deployments, replace with a Redis-backed implementation.
- */
-export declare class RateLimiter {
-    private readonly windowMs;
-    /** Map<userId, sorted array of request timestamps in ms> */
-    private readonly store;
-    constructor(windowMs?: number);
-    /**
-     * Check and record a request for a user.
-     *
-     * @param userId  The user ID (string or numeric)
-     * @param limit   Max allowed requests per window (from aiApiConfig.rateLimitPerMinute)
-     * @returns { allowed: true } or { allowed: false, retryAfterMs: number }
-     */
-    check(userId: string | number, limit: number): {
-        allowed: true;
-    } | {
-        allowed: false;
-        retryAfterMs: number;
-    };
-    /**
-     * Garbage collect entries for inactive users.
-     * Call every ~5 minutes to prevent unbounded memory growth in long-running servers.
-     */
-    gc(): void;
-    /** Clear all state (useful in tests). */
-    clear(): void;
-}

package/dist/server/utils/resolve-service.d.ts

@@ -1,19 +0,0 @@
-import { Context } from '@nocobase/actions';
-/**
- * Resolve an LLM service by name or title.
- */
-export declare function resolveLlmService(ctx: Context, serviceKey: string): Promise<any>;
-/**
- * Resolve a model string to a service + modelId.
- *
- * Strategy (priority order):
- * 1. Try splitting at each "/" position and match the left part against DB (name or title).
- *    This handles cases like "Custom LLM (OpenAI Compatible)/qwen/qwen3.6-plus-preview:free"
- * 2. If no service match, use the defaultLlmService from config and treat the ENTIRE
- *    model string as the modelId. This allows clients to send just "qwen/qwen3.6-plus-preview:free"
- *    or "gpt-4o" without knowing the service name.
- */
-export declare function resolveModelString(ctx: Context, modelString: string): Promise<{
-    service: any;
-    modelId: string;
-} | null>;