EU policy assistant
Step 1 of 21 ยท The system
An employee at a European manufacturer of about 4,000 people, asking how many weeks of parental leave apply in Germany, whether a customer name may go in a ticket, or what the security standard says about USB drives.
An answer that names the file and the section it came from. A fluent paraphrase with no citation is not an answer they can act on.
HR, legal, and security keep their policies as PDFs. Legal edits a policy most weeks. The assistant is only reachable from the company network, and document text, embeddings, and the model call may not leave the EU.
An answer is fluent, plausible, and sourced from a policy that was superseded in February. One of those retires the assistant, because nobody can then tell which of the correct answers were also stale.
An HR employee gets a cited answer from the current HR corpus, a security employee retrieves nothing from the HR-only documents, and a prompt that drops citation quality does not publish.
- 1. The system
- 2. Retrieval, not a fine-tune
- 3. The corpus bucket stays in the EU
- 4. Parse tables and headings before chunking
- 5. A managed knowledge base, not a store you run
- 6. Hierarchical chunking, locked at creation
- 7. One embedding model, matching dimension
- 8. Department metadata, filtered before similarity
- 9. Incremental ingest on weekly edits
- 10. Retrieve, then Converse
- 11. Citations on the answer
- 12. A guardrail, on a published version
- 13. IAM for the model, the profile, and the guardrail
- 14. PrivateLink to bedrock-runtime
- 15. A gateway in front
- 16. The prompt is a versioned template
- 17. Score retrieval and the answer separately
- 18. Invocation logs, encrypted, watched
- 19. The publish gate blocks on an eval drop
- 20. ThrottlingException is a quota question first
- 21. When the right document never comes back
