AI engineering layer · embedding retrieval · measured generation
Steel QC knowledge assistant
A retrieval-augmented assistant over a steel quality-engineering knowledge base. Each passage is a point below. Pick a question to see which passages the system actually read before answering, and how much of the answer those passages support.
How it works
- 1Chunk
Knowledge-base articles are split into short, section-aware passages.
- 2Embed
An embedding model turns each passage into a vector; passages with similar meaning land close together. The vectors live in a Chroma vector database.
- 3Retrieve & answer
The question is embedded the same way, the nearest passages are retrieved, and a language model answers from them with citations.
Where the answer came from
Each dot is a passage, coloured by topic. The map is a UMAP projection of the embedding space. Drag to rotate, scroll to zoom, hover a dot to read it.
Measured, not assumed
Retrieval is scored on questions with a known source passage. Answers are scored for faithfulness: an NLI model checks whether each sentence is entailed by the retrieved passages.