Sample RAG LLM app review
This sample shows the shape of the paid deliverable. A real review is tailored from the buyer's app repo, crawler/source flow, vector-store setup, prompt path, logs, and deployment target.
Example finding
- App state: crawler login, HTML extraction, prompt assembly, model call, local cache files, and compressed API response are coupled into one endpoint, so source-data or credential failures can look like model failures.
- Main risk: requiring all secrets at import time, storing raw source artifacts beside runtime answers, and calling the model without structured retries makes the app hard to operate safely.
- Reliability decision: separate config validation, source authorization, retrieval/context construction, model call, answer cache, and response serialization before exposing the endpoint broadly.
Example next experiments
- Add startup checks for required tokens, object storage, vector index, crawler credentials, and read/write cache paths.
- Record a sanitized retrieval fixture with expected context snippets before tuning prompts.
- Test source-not-found, auth-failed, model-timeout, and cache-hit paths as separate HTTP cases.
- Only then add deployment profiles, observability, and prompt/version tracking.