G1-Proxy — Introduction¶

G1-Proxy is an OpenAI-compatible HTTP proxy that adds real-time AI validation to any LLM backend. It listens for chat requests, screens them, forwards them to the configured upstream model, validates the response, and returns the result — with a single geodesia object attached containing the verdict and the full detection payload.

Call it¶

Point any OpenAI client at the proxy. Nothing else in your code changes — the answer comes back in the shape you already handle, with the verdict attached.

curl -s http://localhost:8080/gw/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"my-model","stream":false,
       "messages":[{"role":"user","content":"What is the capital of France?"}]}' \
  | jq '{answer: .choices[0].message.content, decision: .geodesia.decision, reason: .geodesia.reason}'

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/gw/v1", api_key="not-needed-locally")
r = client.chat.completions.create(
    model="my-model",
    messages=[{"role": "user", "content": "What is the capital of France?"}],
)
print(r.choices[0].message.content)
g = r.model_extra["geodesia"]
print(g["decision"], g["reason"] and g["reason"]["axis"])

import OpenAI from "openai"

const client = new OpenAI({ baseURL: "http://localhost:8080/gw/v1", apiKey: "not-needed-locally" })
const r = (await client.chat.completions.create({
  model: "my-model",
  messages: [{ role: "user", content: "What is the capital of France?" }],
})) as any

console.log(r.choices[0].message.content, r.geodesia.decision, r.geodesia.reason?.axis)

Full request/response contract: Chat API. Every route: Complete API Map.

Key Capabilities¶

Endpoint Summary¶

OpenAI-compatible chat endpoint. Both streaming and non-streaming. Extended with geodesia detection payload.

Ollama-compatible chat endpoint. Mirrors the OpenAI endpoint in behaviour.

Health check. Returns upstream type, logprob capability, and active axis count (4 or 5).

Returns Geodesia's own model descriptor — id, capabilities (e.g. hallucination_detection, tool_calling), and the detected upstream base model — so an OpenAI-compatible client can discover what the deployment supports without a separate probe. See Tool calling.

Returns the current gateway configuration (API key is masked).

Updates one or more configuration fields at runtime. Changes take effect on the next request.

Tests a candidate upstream connection: reachability, model list, logprob support, and a sample reply — all in one call.

Computes black-box token-level causal attribution (occlusion or MuPAX LLM) for a given prompt/response pair.

Serves the built-in user guide as Markdown.

Plus the Knowledge Base (/v1/glad/rag/*) sub-router.

Request & Response Extension¶

Every standard OpenAI request body is accepted verbatim. Geodesia adds the following optional fields:

The full list of request fields (thinking_level, pii_guard, axes, web_search, …) is in the Response Format reference.

The response carries the standard OpenAI structure with exactly one extra top-level key, geodesia (schema 1.0). Its most-read fields:

See API Response Format for the complete object.