1968 passages indexed

Dense RAG desk

Ask a question about anything in the archive — hearings, documents, sources. Every answer links to the exact passages it came from.

Answer

Run a query to see results.

Citations

No citations yet.

Architecture

Dense embeddings (hashed 128-d · MiniLM 384-d). Default offline hashed bi-encoder (concept axes + feature-hashed tokens/char-grams). Optional client MiniLM (Xenova/all-MiniLM-L6-v2) rebuilds the corpus index at 384-d with true neural cosine retrieval.

  1. Query intake
  2. Dense embed (hashed 128-d or MiniLM 384-d)
  3. Cosine top-k retrieve
  4. Sparse hybrid re-rank fusion
  5. Context pack + citation binding
  6. Extractive synthesis or generative Grok (XAI_API_KEY)
Adapters
  • embed: hashed (default) · Xenova MiniLM (client toggle)
  • store: in-memory corpus index (pgvector-ready)
  • generate: extractive · Grok via createServerFn