@almadar/llm

--- package/package.json (excerpt) ---
{
  "name": "@almadar/llm",
  "version": "2.21.0",
  "description": "Multi-provider LLM client with rate limiting, token tracking, structured outputs, and continuation handling",
  "type": "module",
  "main": "./dist/index.js",
  "types": "./dist/index.d.ts",
  "exports": {
    ".": {
      "types": "./dist/index.d.ts",
      "import": "./dist/index.js"
    },
    "./client": {
      "types": "./dist/client.d.ts",
      "import": "./dist/client.js"
    },
    "./json-parser": {
      "types": "./dist/json-parser.d.ts",
      "import": "./dist/json-parser.js"
    },
    "./structured-output": {
      "types": "./dist/structured-output.d.ts",
      "import": "./dist/structured-output.js"
    },
    "./providers": {
      "types": "./dist/providers/index.d.ts",
      "import": "./dist/providers/index.js"
    }
  },
  "files": [
    "dist",
    "src"
  ],
  "dependencies": {
    "@anthropic-ai/sdk": "^0.52.0",
    "@langchain/anthropic": "^1.3.23",
    "@langchain/core": "^1.1.32",
    "@langchain/openai": "^1.2.13",
    "openai": "^6.18.0",
    "zod": "^3.22.0"
  },
  "peerDependencies": {
    "@almadar/core": ">=8.5.0"
  },
  "devDependencies": {
    "@almadar/core": "^8.5.1",
    "@almadar/eslint-plugin": ">=2.3.0",
    "@types/node": "^22.0.0",
    "@typescript-eslint/parser": "8.56.0",
    "eslint": "10.0.0",
    "tsup": "^8.0.0",
    "typescript": "^5.3.0"
  },
  "repository": {
    "type": "git",
    "url": "https://github.com/almadar-io/almadar-llm.git",
    "directory": 

--- package/src/client.ts (excerpt) ---
/**
 * Shared LLM Client
 *
 * Multi-provider LLM client with:
 * - OpenAI, DeepSeek, Anthropic, and Kimi support
 * - Anthropic prompt caching (CachingChatAnthropic)
 * - Rate limiting and retry logic
 * - Token tracking
 * - Structured output parsing with Zod
 *
 * @packageDocumentation
 */

import { ChatOpenAI } from '@langchain/openai';
import { ChatAnthropic } from '@langchain/anthropic';
import type { BaseMessageLike } from '@langchain/core/messages';
import Anthropic from '@anthropic-ai/sdk';
import { z } from 'zod';
import {
  RateLimiter,
  getGlobalRateLimiter,
  type RateLimiterOptions,
} from './rate-limiter.js';
import { TokenTracker, getGlobalTokenTracker } from './token-tracker.js';
import { parseJsonResponse } from './json-parser.js';
import {
  parseChatCompletionResponse,
  type ChatCompletionMessage,
  type ChatCompletionToolDef,
} from './tool-call-types.js';

// ============================================================================
// Local type helpers (avoid Record<string, unknown> and unsafe casts)
// ============================================================================

/** Anthropic generation output with usage metadata (not in Langchain's base types). */
interface AnthropicGenerationWithUsage {
  message?: {
    usage_metadata?: {
      cache_creation_input_tokens?: number;
      cache_read_input_tokens?: number;
      input_tokens?: number;
      output_tokens?: number;
    };
  };
}

/** Response metadata from OpenAI-compatible provid

--- package/src/embedding-client.ts (excerpt) ---
/**
 * Embedding Client
 *
 * Single-purpose client for text embeddings. Used by `@almadar-io/agent`'s
 * cosine-similarity catalog retrieval (rank organisms + atoms against the
 * user's request before rendering Stage A's prompt) and by
 * `@almadar/std`'s publish-time embedding bake step.
 *
 * Providers:
 *   - `openai` (default model `text-embedding-3-small`, 1536-d) — requires
 *     `OPENAI_API_KEY`.
 *   - `openrouter` (default model `baai/bge-base-en-v1.5`, 768-d) — requires
 *     `OPEN_ROUTER_API_KEY`. Same OpenAI-compatible request shape, just a
 *     different base URL.
 *
 * Both providers return the same response envelope (`{data:[{embedding,index}]}`),
 * so the request/response code is shared.
 *
 * @packageDocumentation
 */

export type EmbeddingProvider = 'openai' | 'openrouter';

export interface EmbeddingClientOptions {
  provider: EmbeddingProvider;
  /** Defaults: openai → text-embedding-3-small, openrouter → baai/bge-base-en-v1.5. */
  model?: string;
  /** Override API key. Defaults to provider-specific env var. */
  apiKey?: string;
  /** Override base URL. Defaults to provider canonical endpoint. */
  baseUrl?: string;
}

export interface EmbeddingUsage {
  promptTokens: number;
  totalTokens: number;
}

export interface EmbeddingResult {
  embedding: readonly number[];
  usage: EmbeddingUsage;
}

export interface EmbeddingBatchResult {
  embeddings: readonly (readonly number[])[];
  usage: EmbeddingUsage;
}

interface OpenAIEmbeddingApiResponse {
 

--- dynamic destinations ---
→ api.moonshot.ai (via hostname-var)