Evaluate Endpoint¶

What it does. POST /v1/glad/evaluate on G1-Proxy generates an answer from your upstream LLM and scores it, in one HTTP round-trip. It is the batch-workflow twin of the Chat API: same pipeline, same detection, but you send a bare prompt instead of a message array and you get the full detection payload back rather than a chat-shaped body.

Reach for it when you are scoring a corpus, running an offline evaluation, or wiring detection into something that is not a chat client.

Call it¶

curl -s http://localhost:8080/gw/v1/glad/evaluate \
  -H "Content-Type: application/json" \
  -H "X-Geodesia-App: support_bot" \
  -d '{
    "model": "my-model",
    "prompt": "How tall is the Eiffel Tower?",
    "context": "The Eiffel Tower was built between 1887 and 1889 and stands 330 metres tall."
  }' | jq '{
    answer:   .choices[0].message.content,
    decision: .glad_decision,
    axes:     .geodesia.axis_energy
  }'

import httpx

rows = [
    {"prompt": "How tall is the Eiffel Tower?",
     "context": "The Eiffel Tower … stands 330 metres tall."},
    {"prompt": "Summarise our refund policy.", "context": ""},
]

with httpx.Client(base_url="http://localhost:8080/gw",
                  headers={"X-Geodesia-App": "support_bot"},
                  timeout=120) as c:
    for row in rows:
        r = c.post("/v1/glad/evaluate", json={"model": "my-model", **row}).json()
        axes = r["geodesia"]["axis_energy"]
        flagged = [a for a, e in axes.items() if e.get("flag")]
        print(f"{r['glad_decision']:8s} flagged={flagged or '-'}  {row['prompt'][:40]}")

const rows = [
  { prompt: "How tall is the Eiffel Tower?", context: "…stands 330 metres tall." },
  { prompt: "Summarise our refund policy.", context: "" },
]

for (const row of rows) {
  const r = await fetch("http://localhost:8080/gw/v1/glad/evaluate", {
    method: "POST",
    headers: { "Content-Type": "application/json", "X-Geodesia-App": "support_bot" },
    body: JSON.stringify({ model: "my-model", ...row }),
  }).then(r => r.json())

  const flagged = Object.entries(r.geodesia.axis_energy)
    .filter(([, e]: any) => e.flag)
    .map(([a]) => a)
  console.log(r.glad_decision, flagged, row.prompt.slice(0, 40))
}

What comes back¶

An OpenAI-shaped body carrying the full detection payload — the same structure the Chat API returns:

{
  "choices": [
    { "index": 0,
      "message": { "role": "assistant", "content": "It stands 330 metres tall." },
      "finish_reason": "stop" }
  ],
  "glad_decision": "passed",
  "glad_mode": "blocking",
  "geodesia": {
    "axis_energy": {
      "halluc_context":    { "p_detector": 0.07, "flag": false, "threshold": 0.6475, "available": true },
      "halluc_closedbook": { "p_detector": 0.04, "flag": false, "threshold": 0.58,   "available": true },
      "prompt_safety":     { "p_detector": 0.01, "flag": false, "threshold": 0.9215, "available": true },
      "answer_safety":     { "p_detector": 0.02, "flag": false, "threshold": 0.7295, "available": true },
      "jailbreak":         { "p_detector": 0.00, "flag": false, "threshold": 0.9997, "available": true },
      "rag_jailbreak":     { "p_detector": 0.03, "flag": false, "threshold": 0.2501, "available": true },
      "profanity":         { "p_detector": 0.00, "flag": false, "threshold": 0.90,   "available": true },
      "out_of_scope":      { "p_detector": 0.02, "flag": false, "threshold": 0.90,   "available": true },
      "prompt_complexity": { "p_detector": 0.18, "flag": false, "threshold": 0.50,   "available": true }
    },
    "brake": false,
    "dominant_axis": "prompt_complexity"
  }
}

Never streamed: this endpoint always returns a single JSON body, whatever stream you send.

Request reference¶

¹ Send one of prompt or messages. Everything else on the body that is not a Geodesia control field is forwarded to the upstream as a generation parameter.

Scoring without generating¶

If you already have both the prompt and the answer and only want them scored — replaying stored traffic, evaluating another system's output, testing a threshold change — do not use this endpoint: it will generate a fresh answer. Use the attribution endpoint instead, which scores text you supply:

curl -s http://localhost:8080/gw/v1/glad/causal-explainability/analyze \
  -H "Content-Type: application/json" \
  -d '{
    "prompt":   "How tall is the Eiffel Tower?",
    "response": "It is 1200 metres tall and located in Berlin.",
    "context":  "The Eiffel Tower … stands 330 metres tall.",
    "method":   "dca"
  }'

See Explainability and Causal Explainability.

Studio-local research endpoints¶

G-1 Studio also exposes an older, local-model evaluation surface at /glad/evaluate, /glad/export_audit and /glad/finetune. These are mounted outside /v1, so the unified port on 8080 does not route to them — they are reachable only on Studio's own port (:8199 by default) and they require a research checkpoint loaded in-process. In the packaged product the Studio backend runs without a GPU model, so they are not the path to build on.

For anything production, use POST /v1/glad/evaluate on G1-Proxy — it scores against your real upstream, honours Application policy, and writes to the compliance ledger.