How it works

AIethicsChat is a sources-first research interface backed by retrieval over a curated corpus.

Core idea

1. Your question is embedded into a vector representation.
2. The system runs a nearest-neighbor search over precomputed document chunk embeddings.
3. The top matching chunks are returned as sources.
4. Optionally, an answer is generated that is expected to be grounded in those sources.

Explain mode

Explain mode exposes retrieval diagnostics (process steps, distances, and other metadata) to help:

• understand why a source was retrieved
• debug drift and regressions
• support classroom transparency

Public mode

Public mode is a safer default designed for public deployments. It can clamp or ignore advanced overrides that could:

• reduce relevance (e.g., embedding mismatch)
• increase cost (e.g., unrestricted model overrides)

Knowledge-graph (KG) mode

KG mode can optionally expand retrieval by adding related chunks from a lightweight database-backed graph, then rescoring. If KG data is not available, the system falls back to normal retrieval.

Limitations

• Retrieval quality depends on the corpus: coverage gaps and licensing constraints matter.
• Semantic search can return partially relevant sources; users should inspect sources before relying on claims.
• Generated answers can be wrong; sources are the primary reference.