Trace Format
Trace Format
Each entry in rag_trace.json is a TraceEvent object.
| Field | Type | Description |
|---|---|---|
run_id |
str |
UUID for one run execution. |
corpus_id |
str |
Corpus id derived from input filename stem. |
query |
str |
Query string sent to the orchestrator. |
timestamp_utc |
str |
UTC timestamp for the run event (ISO 8601). |
failure_classes |
list[str] |
Ordered list of detected failures. |
repair_attempts |
int |
Number of executed repair attempts. |
outcome |
str |
Final decision outcome (ACCEPTED, REPAIRED_ACCEPTED, ABSTAINED, FAILED). |
top_score |
float |
Threshold score used for decisioning (BM25 top score). |
chunks_retrieved |
int |
Number of chunks used for generation/verification. |
unsupported_ratio |
float |
Fraction of unsupported response sentences from verifier. |
chunk_ids |
list[str] |
Chunk ids used in the final context. |
raw_top_score |
float |
Raw top retrieval score before any post-processing. |
Example rag_trace.json:
[
{
"run_id": "8c3fc4b8-0c73-4b6b-8c9f-0b6b2bb6d4b7",
"corpus_id": "corpus",
"query": "Where is the Eiffel Tower?",
"timestamp_utc": "2026-05-09T18:30:45.123456+00:00",
"failure_classes": [],
"repair_attempts": 0,
"outcome": "ACCEPTED",
"top_score": 1.2345,
"chunks_retrieved": 3,
"unsupported_ratio": 0.0,
"chunk_ids": ["c1", "c2", "c3"],
"raw_top_score": 1.2345
}
]
The trace file is append-only: multiple runs accumulate in one file.