navra + OpenAI clients

Prerequisites

• navra running with at least one model configured (Ollama, Mistral, or any OpenAI-compatible backend)
• A bearer token

How it works

navra exposes an OpenAI-compatible /v1/chat/completions endpoint. Any client that speaks the OpenAI Chat API can connect -- the client thinks it is talking to OpenAI, but navra routes the request through its safety pipeline and model backend.

Python (openai SDK)

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:9315/v1",
    api_key="mcd_your_token_here",
)

response = client.chat.completions.create(
    model="granite3.3:8b",  # model name from your config.toml
    messages=[
        {"role": "user", "content": "Summarize the key points of IFC."}
    ],
)
print(response.choices[0].message.content)

With tool use

tools = [
    {
        "type": "function",
        "function": {
            "name": "file_read",
            "description": "Read a file from disk",
            "parameters": {
                "type": "object",
                "properties": {
                    "path": {"type": "string"}
                },
                "required": ["path"]
            }
        }
    }
]

response = client.chat.completions.create(
    model="granite3.3:8b",
    messages=[{"role": "user", "content": "Read README.md"}],
    tools=tools,
)

navra enforces ACLs and safety filters on every tool call, even when the request comes through the OpenAI-compatible endpoint.

Node.js (openai SDK)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "http://localhost:9315/v1",
  apiKey: "mcd_your_token_here",
});

const response = await client.chat.completions.create({
  model: "granite3.3:8b",
  messages: [{ role: "user", content: "What is navra?" }],
});

console.log(response.choices[0].message.content);

curl

curl -s http://localhost:9315/v1/chat/completions \
  -H "Authorization: Bearer mcd_your_token_here" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "granite3.3:8b",
    "messages": [{"role": "user", "content": "Hello"}]
  }' | jq .choices[0].message.content

Streaming

All clients support streaming by setting stream=True (Python) or stream: true (Node/curl). navra proxies the SSE stream from the model backend.

Model routing

The model parameter maps to a model name in your config.toml:

[models.granite]
backend = "ollama"
model = "granite3.3:8b"
task = "chat"

If no model name matches, navra returns a 404.

Troubleshooting

"model not found"

Check navra status to see registered models. The model name in the API call must match a key in [models.*] in your config.

Responses are slow

Model inference runs locally or on a remote backend. Check your model backend (Ollama, llama-server) is running and responsive.

Safety filter blocks content

If the response contains [REDACTED:...] markers, navra's safety pipeline detected sensitive content. This is working as intended. Adjust the safety profile if the redaction is too aggressive for your use case.