Triage record: ${INCIDENT_ID}¶

Copy this file for each incident. Fill in one row per failing question. Keep it with the incident. Do not write keys, tokens, real user names or client names. Use IDs and ${PLACEHOLDERS}.

1. Incident¶

Item	Detail
Incident or ticket	${INCIDENT_ID}
Reported by and how
Triage done by
Triage started	YYYY-MM-DD HH:MM
Window of first failure	From YYYY-MM-DD HH:MM to YYYY-MM-DD HH:MM
Index or alias queried	${INDEX_OR_ALIAS}
API version used by the probe
Filter applied for the user	${USER_FILTER}
Filter mode (if the application sets one)

2. Changes in the window¶

List every release, re-index run, alias swap, model change, prompt change, data load or permission change between the last good evaluation and the first failure.

Change	Date and time	Made by	Link or ID

3. Cases¶

One row per failing question. Rank is the position of the expected document in the first N results, or - if it is not there. Copy the probe output into the evidence column.

Case	Question	Time	Expected document id	Keyword rank	Vector rank	Hybrid rank	Hybrid, no filter rank	Class	Evidence
1

Class is one of: retrieval miss, ranking issue, filter issue, generation fault. Every case has exactly one class and one line of evidence.

4. Summary¶

Class	Cases	Likely root cause	Hand over to
Retrieval miss
Ranking issue
Filter issue
Generation fault

5. Severity¶

Item	Detail
Golden set run	Date and run ID
Retrieval metrics against the baseline	recall@10 and nDCG@10, with the tolerance used
Generation metrics against the baseline	Groundedness and relevance, with the tolerance used
Severity	Isolated (a few cases) or systemic (the golden set drops beyond tolerance)

6. Outcome¶

Item	Detail
Root cause
Playbook the incident owner follows
Incident owner
New golden questions to add	Passed to PB-RAG-02-01 on YYYY-MM-DD
Triage closed	YYYY-MM-DD HH:MM