Zero-integration audit

Use historical exports when live SDK integration is unavailable. Current direct formats are Retell, Vapi, Intercom, generic CSV, and generic JSONL.

#Prepare

Create dedicated tenant key so report contains only audit batch. Use customer approved transfer, retention, and deletion process.

Preferred JSONL has one session/line:

{"session_id":"call-42","turns":[{"role":"user","text":"Cancel it"},{"role":"assistant","text":"Which account?"},{"role":"user","text":"No, tomorrow's booking."}]}

#Dry run

python -m importer.replay \
  --input calls.jsonl \
  --gate https://api.intentlayer.dev \
  --tenant-key il_xxx \
  --dry-run

Parser reports malformed/skipped rows without aborting whole batch.

#Run audit

python -m importer.audit \
  --input calls.jsonl \
  --gate https://api.intentlayer.dev \
  --tenant-key il_xxx \
  --out audit/ \
  --redact

Importer detects format, replays turns, waits briefly, downloads report.html, and writes summary.md, checkpoint.json, run.json.

#Observability exports

Langfuse, LangSmith, and OTLP direct parsers are not implemented. Export/transform to generic JSONL or CSV first. Preserve stable conversation ID, turn role/text, and optional timestamp. Local metadata is not sent during replay.

#Read results carefully

Summary estimates correction/rephrase types from transcript text because server report lacks event-type breakdown. Monthly cost estimate is failures × failure_cost × 30 / represented_days; default failure cost is $12. Both are labeled estimates.