gaptrace-evaluate

Scores captured runs. Two tasks, one discovery helper:

pip install gaptrace-evaluate

Is this run healthy? — check()

Call before paying for an LLM; put it in CI.

import gaptrace_capture
from gaptrace_evaluate import check

run_id = gaptrace_capture.capture(
    "what is RRF?", "RRF fuses rankings.",
    chunks=[{"chunk_id": "c1", "source_doc_id": "d1",
             "content": "RRF combines rankings.", "token_count": 10,
             "rerank_score": 0.9}],
)
result = check(run_id)

print(result.verdict)      # "ok" | "warn" | "fail"
print(result.problems)     # ["duplicate chunks: ratio 0.50 exceeds 0.20", ...]
print(result.risk_score)   # 0.0-1.0, None if input quality couldn't be assessed
print(result.factors)      # per-factor {value, threshold, status}
print(result.thresholds)   # "learned" | "policy" — which standards were applied

check() compares all free input metrics against the current standards: once at least 10 evaluated runs exist for the pipeline it uses thresholds learned from your own history (and says so via thresholds == "learned"); before that it falls back to the policy defaults. A run captured without chunks gets a warn verdict explaining the missing data — never an exception.

Score it fully — evaluate()

import gaptrace_capture
from gaptrace_evaluate import evaluate

run_id = gaptrace_capture.capture(
    "what is RRF?", "RRF fuses rankings.",
    chunks=[{"chunk_id": "c1", "source_doc_id": "d1",
             "content": "RRF combines rankings.", "token_count": 10,
             "rerank_score": 0.9}],
)

# Complete eval: every metric applicable to the record.
result = evaluate(run_id)

# One atomic metric — nothing else is computed:
result = evaluate(run_id, metrics=["duplicates"], save=False)

# A chosen subset:
result = evaluate(run_id, metrics=["relevance", "faithfulness"])

target can be an sNrN string (what gaptrace_capture.capture() returns), a committed Capture object, or a bare RunRecord (then pass save=False — there's no run row to write to).

EvalResult

save=True (default) persists via the one store path; gaptrace explain <run_id> then shows the scores alongside its analysis.

available_metrics()

Output metrics are RAGAS LLM-as-judge calls — they cost money and need a configured judge. To stay free-only, use check(), or select input metrics explicitly: evaluate(run_id, metrics=["relevance", "duplicates", "truncation", "token_efficiency", "coherence"]).

Policy system

Human-set thresholds encoding known failure modes; active from day one and the fallback standard for check(). Stored per pipeline.

gaptrace-evaluate policy show
gaptrace-evaluate policy set max_duplicate_ratio 0.1
gaptrace-evaluate policy reset

Programmatic override: evaluate(run_id, policy=InputQualityPolicy(...)).

Benchmark lifecycle (CLI)

Learned thresholds accumulate as you evaluate real runs — check() picks them up automatically at 10+ evaluated runs. The CLI exposes the machinery for inspection:

gaptrace-evaluate run s2r3                 # evaluate one run (both layers)
gaptrace-evaluate run s2r3 --input-only    # free metrics only
gaptrace-evaluate run --session s2         # evaluate a whole session
gaptrace-evaluate benchmark show           # current learned thresholds
gaptrace-evaluate benchmark build          # rebuild from evaluated history
gaptrace-evaluate benchmark check s2r3     # factor-by-factor threshold check
gaptrace-evaluate benchmark export         # RAGAS-compatible JSONL dataset

--input-only --output-only together is an error (it would compute nothing).