Work / AI engineering demo
Corpus
Retrieval that can show its work.
Query analysis, retrieval, ranked context, then an extract or a refusal. Full-text search always runs. Dense search is added only when embeddings are stored.
- Problem
- A fluent answer without a retrieved source is not evidence.
- Role
- Independent engineer. Public demo on this site, not an employer case study.
- Implemented scope
- Chunk the public catalog, rank passages, keep a score floor, and attach citations or refuse.
Architecture
- 01
Query analysis
The question is read as a retrieval query, not as a license to invent an answer.
- Responsibility
- Frame the question
- Concern
- Stay inside the catalog
- 02
Retrieval
Full-text search always runs. Dense search and fusion run when embeddings are stored.
- Responsibility
- Rank catalog passages
- Concern
- Score floor
- 03
Context
Ranked, attributed chunks are the only context passed forward.
- Responsibility
- Assemble evidence
- Concern
- Provenance
- 04
Answer or refuse
The response cites a source or refuses. Low similarity does not become a fluent guess.
- Responsibility
- Cite or stop
- Concern
- No unsupported answer
What holds
- Decision
- Keep lexical search as the always-on path. Add dense retrieval only when vectors exist.
- Failure handling
- Low similarity does not reach generation. The answer refuses instead of guessing.
- Evidence
- The lab walkthrough runs query, retrieve, assemble, then extract or refuse. It is labeled as an interactive demo, not an accuracy score.
- Status
- AI engineering demo
- Limits
- No eval scores are published. Chat uses a model only when that request succeeds; otherwise the path is a local extract.