# LLM Inference API Documentation

OpenAI-compatible REST API for unified LLM inference across multiple backends.

## Base URL

```
{BASE_URL}
```

Common configurations:
- **Local development:** `http://localhost:14011`
- **Docker (internal):** `http://didiAI-llm-api:14011`
- **Docker (external):** `http://<host>:14011`
- **VPN/Production:** Use your configured hostname or IP

## Authentication

### Bearer Token Authentication (Optional)

The API supports optional Bearer token authentication. When enabled, protected endpoints require a valid token.

**Enable by setting:**
```bash
LLM_API_TOKENS=token1,token2,token3  # comma-separated for multiple tokens
```

**Request header:**
```
Authorization: Bearer <your-token>
```

**Public endpoints (no auth required):**
- `GET /health`
- `GET /ready`

**Protected endpoints (require auth when enabled):**
- `POST /v1/chat/completions`
- `POST /v1/completions`
- `GET /v1/models`
- `POST /v1/models/load`
- `POST /v1/models/unload`
- `GET /v1/backends`

**Error response (401):**
```json
{
  "detail": {
    "error": "Authentication required",
    "message": "Missing Authorization header"
  }
}
```

Response includes `WWW-Authenticate: Bearer` header.

### Backend API Keys

Backend services require their own API keys:
- **LiteLLM**: `OPENROUTER_API_KEY`, `OPENAI_API_KEY`, or `ANTHROPIC_API_KEY`
- **vLLM**: Optional `LLM_VLLM_API_KEY` for vLLM server

> **Reasoning model note (`LLM_VLLM_DISABLE_THINKING`, default `true`):** the local model `Qwen/Qwen3.5-35B-A3B` (served as `qwen3.5`) is a reasoning model. By default the gateway injects `chat_template_kwargs={"enable_thinking": false}` on vLLM chat requests so the model returns the final answer directly instead of a `thinking` preamble — important for callers that parse JSON. Callers may override by passing their own `chat_template_kwargs` in the request body.

## Rate Limiting

The API uses token bucket rate limiting:

| Header | Description |
|--------|-------------|
| `Retry-After` | Seconds to wait when rate limited (429 response) |
| `X-Request-ID` | Unique request identifier (auto-generated or pass via header) |

Default limits:
- **10 requests/second** with burst of 20
- **10 concurrent completion requests**

Configure via environment variables:
- `LLM_RATE_LIMIT_RPS`: Requests per second
- `LLM_RATE_LIMIT_BURST`: Maximum burst size
- `LLM_MAX_CONCURRENT_COMPLETIONS`: Concurrent request limit

---

## Endpoints

### Chat Completions

Create a chat completion with optional streaming.

```
POST /v1/chat/completions
```

#### Request Body

| Field | Type | Required | Default | Description |
|-------|------|----------|---------|-------------|
| `messages` | array | Yes | - | List of messages (1-1000) |
| `model` | string | Yes | - | Model identifier |
| `temperature` | float | No | 0.7 | Sampling temperature (0.0-2.0) |
| `max_tokens` | integer | No | null | Maximum tokens to generate (1-1,000,000) |
| `stream` | boolean | No | false | Enable streaming response |
| `backend` | string | No | null | Backend override: `litellm`, `vllm`, `llamacpp` |
| `top_p` | float | No | null | Top-p sampling (0.0-1.0) |
| `frequency_penalty` | float | No | null | Frequency penalty (-2.0 to 2.0) |
| `presence_penalty` | float | No | null | Presence penalty (-2.0 to 2.0) |
| `stop` | string/array | No | null | Stop sequences |

#### Message Object

| Field | Type | Required | Description |
|-------|------|----------|-------------|
| `role` | string | Yes | One of: `system`, `user`, `assistant`, `function`, `tool` |
| `content` | string | Yes* | Message content (*can be null for assistant role) |
| `name` | string | No | Optional author name |

#### Response (Non-Streaming)

```json
{
  "id": "chatcmpl-abc123",
  "object": "chat.completion",
  "created": 1704067200,
  "model": "gpt-3.5-turbo",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Hello! How can I help you today?"
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 10,
    "completion_tokens": 15,
    "total_tokens": 25
  },
  "backend": "litellm"
}
```

#### Response (Streaming)

Server-Sent Events (SSE) format. Each event contains a JSON chunk:

```
data: {"id":"chatcmpl-abc123","object":"chat.completion.chunk","created":1704067200,"model":"gpt-3.5-turbo","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

data: {"id":"chatcmpl-abc123","object":"chat.completion.chunk","created":1704067200,"model":"gpt-3.5-turbo","choices":[{"index":0,"delta":{"content":"Hello"},"finish_reason":null}]}

data: {"id":"chatcmpl-abc123","object":"chat.completion.chunk","created":1704067200,"model":"gpt-3.5-turbo","choices":[{"index":0,"delta":{"content":"!"},"finish_reason":null}]}

data: {"id":"chatcmpl-abc123","object":"chat.completion.chunk","created":1704067200,"model":"gpt-3.5-turbo","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

data: [DONE]
```

#### Example Request

```bash
# Non-streaming
curl -X POST http://localhost:14011/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "What is the capital of France?"}
    ],
    "model": "gpt-3.5-turbo",
    "temperature": 0.7,
    "max_tokens": 100
  }'

# Streaming
curl -X POST http://localhost:14011/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{
    "messages": [{"role": "user", "content": "Tell me a short story"}],
    "model": "gpt-4",
    "stream": true
  }'

# With specific backend
curl -X POST http://localhost:14011/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "Hello!"}],
    "model": "meta-llama/Llama-2-7b-chat-hf",
    "backend": "vllm"
  }'
```

---

### Text Completions (legacy)

Create a non-streaming text completion from a raw prompt. OpenAI-compatible legacy `/v1/completions`. Routes to the model's backend (vLLM primary, cloud via LiteLLM). Backends that do not support text completion return `501`.

```
POST /v1/completions
```

#### Request Body

| Field | Type | Required | Default | Description |
|-------|------|----------|---------|-------------|
| `prompt` | string/array | Yes | - | Prompt(s) to complete |
| `model` | string | No | server default | Model (alias) to use |
| `temperature` | float | No | 0.7 | Sampling temperature (0.0-2.0) |
| `max_tokens` | integer | No | null | Maximum tokens to generate (1-1,000,000) |
| `backend` | string | No | null | Backend override: `litellm`, `vllm`, `llamacpp` |
| `top_p` | float | No | null | Top-p sampling (0.0-1.0) |
| `frequency_penalty` | float | No | null | Frequency penalty (-2.0 to 2.0) |
| `presence_penalty` | float | No | null | Presence penalty (-2.0 to 2.0) |
| `stop` | string/array | No | null | Stop sequences |

#### Response

```json
{
  "id": "cmpl-abc123",
  "object": "text_completion",
  "created": 1704067200,
  "model": "qwen3.5",
  "choices": [
    {
      "index": 0,
      "text": "Paris is the capital of France.",
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 8,
    "completion_tokens": 7,
    "total_tokens": 15
  },
  "backend": "vllm"
}
```

#### Example Request

```bash
curl -X POST http://localhost:14011/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "The capital of France is",
    "model": "qwen3.5",
    "max_tokens": 16
  }'
```

---

### List Models

List available models across all or specific backends.

```
GET /v1/models
GET /v1/models?backend=litellm
```

#### Query Parameters

| Parameter | Type | Required | Description |
|-----------|------|----------|-------------|
| `backend` | string | No | Filter by backend: `litellm`, `vllm`, `llamacpp` |

#### Response

```json
{
  "object": "list",
  "data": [
    {
      "id": "gpt-3.5-turbo",
      "backend": "litellm",
      "loaded": true,
      "context_length": 16384,
      "capabilities": ["chat"]
    },
    {
      "id": "gpt-4",
      "backend": "litellm",
      "loaded": true,
      "context_length": 128000,
      "capabilities": ["chat"]
    },
    {
      "id": "meta-llama/Llama-2-7b-chat-hf",
      "backend": "vllm",
      "loaded": true,
      "context_length": 4096,
      "capabilities": ["chat"]
    }
  ]
}
```

#### Example Request

```bash
# All models
curl http://localhost:14011/v1/models

# Models from specific backend
curl "http://localhost:14011/v1/models?backend=vllm"
```

---

### Load Model

Load a model on a local backend (vLLM or llama.cpp only).

```
POST /v1/models/load
```

#### Request Body

| Field | Type | Required | Description |
|-------|------|----------|-------------|
| `model` | string | Yes | Model identifier to load |
| `backend` | string | Yes | Target backend: `vllm` or `llamacpp` |

#### Response

```json
{
  "success": true,
  "model": "meta-llama/Llama-2-7b-chat-hf",
  "backend": "vllm",
  "message": "Model loaded successfully"
}
```

#### Example Request

```bash
curl -X POST http://localhost:14011/v1/models/load \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta-llama/Llama-2-7b-chat-hf",
    "backend": "vllm"
  }'
```

---

### Unload Model

Unload a model from a local backend (vLLM or llama.cpp only).

```
POST /v1/models/unload
```

#### Request Body

| Field | Type | Required | Description |
|-------|------|----------|-------------|
| `model` | string | Yes | Model identifier to unload |
| `backend` | string | Yes | Backend to unload from: `vllm` or `llamacpp` |

#### Response

```json
{
  "success": true,
  "model": "meta-llama/Llama-2-7b-chat-hf",
  "backend": "vllm",
  "message": "Model unloaded successfully"
}
```

#### Example Request

```bash
curl -X POST http://localhost:14011/v1/models/unload \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta-llama/Llama-2-7b-chat-hf",
    "backend": "vllm"
  }'
```

---

### List Backends

List available backends.

```
GET /v1/backends
```

#### Response

```json
{
  "backends": ["litellm", "vllm", "llamacpp"]
}
```

#### Example Request

```bash
curl http://localhost:14011/v1/backends
```

---

### Health Check

Check overall system health and per-backend status.

```
GET /health
```

#### Response

```json
{
  "status": "healthy",
  "backends": [
    {
      "name": "litellm",
      "healthy": true,
      "message": null
    },
    {
      "name": "vllm",
      "healthy": true,
      "message": null
    },
    {
      "name": "llamacpp",
      "healthy": false,
      "message": "Connection refused"
    }
  ]
}
```

#### Status Values

| Status | Description |
|--------|-------------|
| `healthy` | All backends are healthy |
| `degraded` | Some backends are healthy |
| `unhealthy` | No backends are healthy |

#### Example Request

```bash
curl http://localhost:14011/health
```

---

### Readiness Probe

Kubernetes-style readiness probe. Returns ready if the default backend is healthy.

```
GET /ready
```

#### Response

```json
{
  "ready": true
}
```

#### HTTP Status Codes

| Code | Description |
|------|-------------|
| 200 | Service is ready |
| 503 | Service is not ready |

#### Example Request

```bash
curl http://localhost:14011/ready
```

---

### Service Info (catalog)

Service metadata for cross-module catalog integration. Returns `resource`, `models` (from all enabled backends), and `functions` (API endpoint descriptors). Consumed by the catalog-api.

```
GET /v1/info
```

#### Example Request

```bash
curl http://localhost:14011/v1/info
```

---

## Error Responses

All errors follow a consistent format:

```json
{
  "detail": "Error message description"
}
```

### HTTP Status Codes

| Code | Description |
|------|-------------|
| 400 | Bad Request - Invalid parameters or backend |
| 401 | Unauthorized - Missing or invalid Bearer token (when auth enabled) |
| 422 | Validation Error - Request body validation failed |
| 429 | Too Many Requests - Rate limit exceeded |
| 500 | Internal Server Error - Completion or backend failure |
| 503 | Service Unavailable - Concurrency limit exceeded or service not ready |

### Error Examples

**Validation Error (422)**
```json
{
  "detail": [
    {
      "type": "value_error",
      "loc": ["body", "messages", 0, "content"],
      "msg": "Message at index 0 with role 'user' cannot have empty content",
      "input": ""
    }
  ]
}
```

**Rate Limit (429)**
```json
{
  "detail": "Rate limit exceeded"
}
```
Response includes `Retry-After` header with seconds to wait.

**Backend Error (400)**
```json
{
  "detail": "Backend 'vllm' is not available"
}
```

**Concurrency Limit (503)**
```json
{
  "detail": "Concurrency limit exceeded"
}
```

---

## Request Headers

| Header | Required | Description |
|--------|----------|-------------|
| `Content-Type` | Yes (POST) | Must be `application/json` |
| `Authorization` | When auth enabled | Bearer token: `Bearer <your-token>` |
| `Accept` | No | Use `text/event-stream` for streaming |
| `X-Request-ID` | No | Custom request ID (auto-generated if not provided) |

---

## Supported Backends

### LiteLLM (cloud)

Supports 100+ LLM providers through a unified interface. (`LLM_DEFAULT_BACKEND` is required and has no default; the DIDI deployment runs `vllm` with the local `qwen3.5` model and uses LiteLLM for cloud/premium models.)

**Popular models:**
- `gpt-3.5-turbo`, `gpt-4`, `gpt-4-turbo` (OpenAI)
- `claude-3-opus`, `claude-3-sonnet`, `claude-3-haiku` (Anthropic)
- `openrouter/meta-llama/llama-3-70b` (OpenRouter)

### vLLM

High-throughput GPU inference for open-source models.

**Requirements:** NVIDIA GPU with 16GB+ VRAM

**Example models:**
- `meta-llama/Llama-2-7b-chat-hf`
- `mistralai/Mistral-7B-Instruct-v0.2`
- `microsoft/phi-2`

### llama.cpp

CPU/Metal inference for GGUF format models.

**Requirements:** 16GB+ RAM, CPU with AVX2 support

**Example models:** Local GGUF files

---

## Python Client Example

```python
import httpx

async def chat_completion():
    async with httpx.AsyncClient() as client:
        response = await client.post(
            "http://localhost:14011/v1/chat/completions",
            json={
                "messages": [{"role": "user", "content": "Hello!"}],
                "model": "gpt-3.5-turbo",
            },
        )
        return response.json()

async def streaming_completion():
    async with httpx.AsyncClient() as client:
        async with client.stream(
            "POST",
            "http://localhost:14011/v1/chat/completions",
            json={
                "messages": [{"role": "user", "content": "Tell me a story"}],
                "model": "gpt-4",
                "stream": True,
            },
        ) as response:
            async for line in response.aiter_lines():
                if line.startswith("data: "):
                    data = line[6:]
                    if data != "[DONE]":
                        chunk = json.loads(data)
                        content = chunk["choices"][0]["delta"].get("content", "")
                        print(content, end="", flush=True)
```

---

## OpenAI SDK Compatibility

The API is compatible with the OpenAI Python SDK:

```python
from openai import OpenAI

# Without auth (when LLM_API_TOKENS not set)
client = OpenAI(
    base_url="http://localhost:14011/v1",
    api_key="not-needed",  # Required by SDK but not used
)

# With auth (when LLM_API_TOKENS is set)
client = OpenAI(
    base_url="http://localhost:14011/v1",
    api_key="your-api-token",  # Your LLM_API_TOKENS value
)

response = client.chat.completions.create(
    model="gpt-3.5-turbo",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)

# Streaming
for chunk in client.chat.completions.create(
    model="gpt-4",
    messages=[{"role": "user", "content": "Tell me a story"}],
    stream=True,
):
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")
```