Token-level observability
Logprobs + top tokens
Every token returned by Meridian carries a log-probability score. Enable logprobs=true and set top_logprobs=5 to surface the five most likely alternatives alongside their scores.
Quick start
curl https://api.getnimbus.net/v1/chat/completions \
-H "Authorization: Bearer $MERIDIAN_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "meridian-1",
"messages": [{"role":"user","content":"Explain entropy"}],
"logprobs": true,
"top_logprobs": 5
}'The response payload includes a logprobs array on every choice, with per-token scores and the top-5 alternates.
Confidence scoring
Aggregate token logprobs into a sequence-level confidence metric. Low-confidence spans flag hallucinations, ambiguous completions, or out-of-distribution inputs — ideal for guardrails and human-in-the-loop workflows.
- Mean logprob — average over all generated tokens; simple threshold gating.
- Min logprob — catch the single least-certain token in a response.
- Entropy gap — distance between the chosen token and the runner-up; narrow gaps signal indecision.
RAG re-ranking
When Meridian generates an answer grounded in retrieved chunks, logprobs reveal which tokens the model is most certain about. Re-rank candidate passages by the cumulative logprob of the tokens they contributed — higher fidelity than embedding cosine alone.
// Pseudo: re-rank chunks by token-level contribution
const scores = chunks.map((chunk, i) => {
const tokens = tokenize(chunk.text);
const lp = tokens.reduce((sum, t) => sum + (t.logprob ?? 0), 0);
return { index: i, score: lp / tokens.length };
});
scores.sort((a, b) => b.score - a.score);Response shape
{
"choices": [{
"logprobs": {
"content": [
{
"token": " Ent",
"logprob": -0.12,
"top_logprobs": [
{ "token": " Ent", "logprob": -0.12 },
{ "token": " The", "logprob": -2.41 },
{ "token": " In", "logprob": -3.10 },
{ "token": " A", "logprob": -4.05 },
{ "token": " Sh", "logprob": -5.20 }
]
}
]
}
}]
}Streaming responses emit logprobs incrementally on each chunk — no buffering required.
Ready to calibrate your outputs?
Grab an API key and start sending logprobs=true today.