Skip to content
AI-Lib

Provider Manifests

Provider Manifests

Section titled “Provider Manifests”

Each AI provider in the ecosystem has a YAML manifest file (v1/providers/<provider>.yaml) that fully describes how to interact with its API.

Supported Providers

Section titled “Supported Providers”

Global Providers

Section titled “Global Providers”

OpenAI, Anthropic, Google Gemini, Groq, Mistral, Cohere, Perplexity, Together AI, DeepInfra, OpenRouter, Azure OpenAI, NVIDIA, Fireworks AI, Replicate, AI21 Labs, Cerebras, Lepton AI, Grok

China Region Providers

Section titled “China Region Providers”

DeepSeek, Qwen (Alibaba), Zhipu GLM, Doubao (ByteDance), Baidu ERNIE, iFlytek Spark, Tencent Hunyuan, SenseNova, Tiangong, Moonshot (Kimi), MiniMax, Baichuan, Yi (01.AI), SiliconFlow

Manifest Structure

Section titled “Manifest Structure”

Endpoint Configuration

Section titled “Endpoint Configuration”
endpoint:
  base_url: "https://api.openai.com/v1"
  chat_path: "/chat/completions"
  protocol: "https"
  timeout_ms: 60000

Authentication

Section titled “Authentication”

Supports multiple auth types:

# Bearer token (most common)
auth:
  type: bearer
  token_env: "OPENAI_API_KEY"


# API key in header
auth:
  type: api_key
  header: "x-api-key"
  token_env: "ANTHROPIC_API_KEY"


# Custom headers
auth:
  type: bearer
  token_env: "ANTHROPIC_API_KEY"
  headers:
    anthropic-version: "2023-06-01"

Parameter Mappings

Section titled “Parameter Mappings”

Maps standard parameter names to provider-specific fields:

parameter_mappings:
  temperature: "temperature"
  max_tokens: "max_completion_tokens"  # OpenAI uses different name
  stream: "stream"
  tools: "tools"
  tool_choice: "tool_choice"
  response_format: "response_format"

Streaming Configuration

Section titled “Streaming Configuration”

Declares how to decode and interpret streaming responses:

streaming:
  decoder:
    format: "sse"              # "sse", "ndjson", or "anthropic_sse"
    done_signal: "[DONE]"      # Stream termination marker
  event_map:
    - match: "$.choices[0].delta.content"
      emit: "PartialContentDelta"
      extract:
        content: "$.choices[0].delta.content"
    - match: "$.choices[0].delta.tool_calls"
      emit: "PartialToolCall"
      extract:
        tool_calls: "$.choices[0].delta.tool_calls"
    - match: "$.choices[0].finish_reason"
      emit: "StreamEnd"
      extract:
        finish_reason: "$.choices[0].finish_reason"

Error Classification

Section titled “Error Classification”

Maps HTTP responses to standard error types:

error_classification:
  by_http_status:
    "400": "invalid_request"
    "401": "authentication"
    "403": "permission"
    "404": "not_found"
    "429": "rate_limited"
    "500": "server_error"
    "503": "overloaded"
  by_error_code:
    "context_length_exceeded": "context_length"
    "content_filter": "content_filter"

Capabilities

Section titled “Capabilities”

Feature flags that runtimes check before making requests:

capabilities:
  streaming: true
  tools: true
  vision: true
  audio: false
  reasoning: true
  agentic: true
  json_mode: true

How Runtimes Use Manifests

Section titled “How Runtimes Use Manifests”
  1. Load — Read YAML manifest (local, env var, or GitHub)
  2. Validate — Check against JSON Schema
  3. Compile — Convert user request using parameter mappings
  4. Execute — Send HTTP request with correct auth/headers
  5. Decode — Process response using streaming configuration
  6. Classify — Handle errors using classification rules

Next Steps

Section titled “Next Steps”