> ## Documentation Index
> Fetch the complete documentation index at: https://developers.lighton.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# OVHcloud AI Endpoints

> Use LightOn search as the retrieval layer for models hosted on OVHcloud AI Endpoints.

OVHcloud AI Endpoints expose hosted open models (Qwen, Mistral, Llama, gpt-oss, and others) through an OpenAI-compatible endpoint, served from OVHcloud's European infrastructure. Pair them with LightOn search to build RAG pipelines where the retrieval stays on LightOn's infrastructure and the generation runs on OVHcloud.

The flow is:

1. Search LightOn for the passages most relevant to the user's question.
2. Pack those passages into the model's context window.
3. Call the OVHcloud model to generate an answer grounded in the retrieved content.

## Prerequisites

* A `LIGHTON_API_KEY`, available in the **Console → API Keys** section.
* An OVHcloud AI Endpoints access token, available in the OVHcloud Control Panel under **Public Cloud → AI Endpoints**. Store it as `OVH_AI_ENDPOINTS_ACCESS_TOKEN`.
* At least one workspace with indexed documents on LightOn.

## Installation

```bash theme={null}
pip install requests openai
```

The `openai` package is used here only for its client; OVHcloud's `/chat/completions` endpoint is fully compatible with it.

## Full example

```python theme={null}
import os
import requests
from openai import OpenAI

LIGHTON_API_KEY = os.environ["LIGHTON_API_KEY"]
OVH_AI_ENDPOINTS_ACCESS_TOKEN = os.environ["OVH_AI_ENDPOINTS_ACCESS_TOKEN"]

ovh = OpenAI(
    base_url="https://oai.endpoints.kepler.ai.cloud.ovh.net/v1",
    api_key=OVH_AI_ENDPOINTS_ACCESS_TOKEN,
)


def search(query: str, workspace_id: list[int] | None = None, max_results: int = 5) -> list[dict]:
    payload = {"query": query, "max_results": max_results}
    if workspace_id:
        payload["workspace_id"] = workspace_id

    response = requests.post(
        "https://api.lighton.ai/api/v3/search",
        headers={"Authorization": f"Bearer {LIGHTON_API_KEY}"},
        json=payload,
    )
    response.raise_for_status()
    return response.json()["results"]


def answer(question: str, workspace_id: list[int] | None = None, model: str = "Meta-Llama-3_3-70B-Instruct") -> str:
    results = search(question, workspace_id=workspace_id)

    context = "\n\n".join(
        f"[{r['source']['filename']}, p.{r['source']['page_start']}]\n{r['content']}"
        for r in results
        if r["content"]
    )

    completion = ovh.chat.completions.create(
        model=model,
        messages=[
            {
                "role": "system",
                "content": (
                    "You are a helpful assistant. Answer the user's question using only "
                    "the provided context. If the context does not contain enough information, "
                    "say so.\n\nContext:\n" + context
                ),
            },
            {"role": "user", "content": question},
        ],
    )
    return completion.choices[0].message.content


print(answer("What is our data retention policy?"))
```

## LightOn search as a tool

Instead of always searching before calling the model, you can expose LightOn search as a tool and let the model decide when to call it. The model issues a `lighton_search` tool call when it needs context; your code executes the search and feeds the results back; the model then produces a final answer.

```python theme={null}
import json
import os
import requests
from openai import OpenAI

LIGHTON_API_KEY = os.environ["LIGHTON_API_KEY"]
OVH_AI_ENDPOINTS_ACCESS_TOKEN = os.environ["OVH_AI_ENDPOINTS_ACCESS_TOKEN"]

ovh = OpenAI(
    base_url="https://oai.endpoints.kepler.ai.cloud.ovh.net/v1",
    api_key=OVH_AI_ENDPOINTS_ACCESS_TOKEN,
)

SEARCH_TOOL = {
    "type": "function",
    "function": {
        "name": "lighton_search",
        "description": (
            "Search the company knowledge base for passages relevant to a query. "
            "Returns ranked excerpts with their source filename and page numbers."
        ),
        "parameters": {
            "type": "object",
            "properties": {
                "query": {
                    "type": "string",
                    "description": "Natural-language search query.",
                },
                "max_results": {
                    "type": "integer",
                    "description": "Number of passages to return (1–50, default 5).",
                    "default": 5,
                },
            },
            "required": ["query"],
        },
    },
}


def run_search(query: str, max_results: int = 5) -> str:
    response = requests.post(
        "https://api.lighton.ai/api/v3/search",
        headers={"Authorization": f"Bearer {LIGHTON_API_KEY}"},
        json={"query": query, "max_results": max_results},
    )
    response.raise_for_status()
    results = response.json()["results"]
    passages = [
        f"[{r['source']['filename']}, p.{r['source']['page_start']}]\n{r['content']}"
        for r in results
        if r["content"]
    ]
    return "\n\n".join(passages) if passages else "No results found."


def answer(question: str, model: str = "Meta-Llama-3_3-70B-Instruct") -> str:
    messages = [{"role": "user", "content": question}]

    while True:
        completion = ovh.chat.completions.create(
            model=model,
            tools=[SEARCH_TOOL],
            messages=messages,
        )
        choice = completion.choices[0]

        if choice.finish_reason == "tool_calls":
            messages.append(choice.message)
            for call in choice.message.tool_calls:
                args = json.loads(call.function.arguments)
                result = run_search(**args)
                messages.append({
                    "role": "tool",
                    "tool_call_id": call.id,
                    "content": result,
                })
        else:
            return choice.message.content


print(answer("What is our data retention policy?"))
```

The loop handles the case where the model issues multiple search calls in sequence before producing a final answer.

## Scoping retrieval to a workspace

Pass `workspace_id` to limit search to a specific workspace. This is useful in multi-tenant products where each customer's data lives in a dedicated workspace.

```python theme={null}
answer("Summarize the onboarding checklist", workspace_id=[42])
```

## Choosing a model

OVHcloud's catalog includes several hosted models. Pass the model name to the `model` parameter:

| Model                                 | Notes                                            |
| ------------------------------------- | ------------------------------------------------ |
| `Meta-Llama-3_3-70B-Instruct`         | Strong reasoning, good default choice            |
| `Llama-3.1-8B-Instruct`               | Faster and cheaper, suitable for simpler queries |
| `Mistral-Small-3.2-24B-Instruct-2506` | Compact Mistral, low latency                     |
| `Qwen3-32B`                           | Strong multilingual reasoning model              |
| `Qwen2.5-VL-72B-Instruct`             | Vision-language model, accepts image input       |

You can list the models available to your token at any time:

```python theme={null}
print([m.id for m in ovh.models.list().data])
```

Check the [OVHcloud AI Endpoints documentation](https://endpoints.ai.cloud.ovh.net/) for the current model list and regional availability.

## Streaming responses

OVHcloud's endpoint supports streaming. Enable it by passing `stream=True` and iterating over the response:

```python theme={null}
stream = ovh.chat.completions.create(
    model="Meta-Llama-3_3-70B-Instruct",
    messages=[...],
    stream=True,
)

for chunk in stream:
    delta = chunk.choices[0].delta.content
    if delta:
        print(delta, end="", flush=True)
```

## A note on reasoning models

Some OVHcloud models (for example `Qwen3-32B` and `Qwen3.6-27B`) are reasoning models. When called through the raw HTTP API they may return their chain of thought under a `reasoning` field and the final answer under `content`. The `openai` client used in the examples above surfaces the final answer in `choices[0].message.content` as usual, so no special handling is needed; read `message.reasoning` only if you want to inspect the thinking trace.