Recipe: Stack trace diagnoser
Build a self-serve tool that ingests a raw stack trace and returns actionable root-cause hypotheses — no human triage required.
Ingredients
- Raw stack trace text (copy-paste from any log source)
- Symbol map or PDB-derived function-name index
- Error-signature corpus (known crash fingerprints)
- LLM prompt template tuned for triage classification
Steps
- Parse the trace — extract module names, offsets, and exception codes via regex.
- Resolve symbols — map each offset to the closest known function using the symbol index.
- Fingerprint — hash the resolved call chain and compare against the error-signature corpus.
- Classify — feed the trace, resolved symbols, and closest fingerprint matches into the LLM prompt. Request a structured JSON response with root-cause category, confidence score, and suggested remediation.
- Render — display the hypothesis card with expandable raw trace, symbol mapping, and a one-click copy for the engineering team.
Pro tips
- Cache symbol resolutions in a local SQLite DB to keep latency under 200 ms.
- Strip PII and hostnames from traces before sending to any external LLM endpoint.
- Maintain a feedback loop — let engineers mark hypotheses as correct/incorrect to retrain the fingerprint corpus.