Skip to content
Concepts
Hybrid RAG

Hybrid RAG

RAG (Retrieval-Augmented Generation) is the mechanism by which Vectora answers based on what was indexed, not on what the model “thinks” is true. It’s a central pillar of the architecture, not a secondary feature — there’s a dedicated subagent (search) and a set of tools (vector_search, embedding, ingest_docs, manage_retriever) built around it.

Pipeline 

user query
      │
      ▼
 query expansion (LLM multi-query)
      │
      ▼
 hybrid search:  BM25 (sparse/lexical)  +  dense vector search (LanceDB/Qdrant)
      │
      ▼
 score gate
      │
   ┌──┴──────────────┬─────────────────────┐
   │ score ≥ 0.7      │ 0.4 ≤ score < 0.7   │ score < 0.4
   ▼                  ▼                     ▼
 inject directly   rerank (Cohere/         falls back to web search (Tavily),
 into context       VoyageAI) → inject      result curated before indexing
  • Hybrid retrieval combines BM25 (good for exact terms, function names, identifiers) with dense vector search (good for semantic similarity) — neither alone covers both cases well.
  • Reranking (Cohere rerank-multilingual-v3.0 or VoyageAI) reorders candidates by actual relevance before they enter context, preventing semantic noise from surfacing the wrong results.
  • Web search fallback: when the local retrieval score is low, Vectora doesn’t force a bad answer — it searches the web (Tavily) and curates the result (reranker + LLM judge) before ever considering indexing it, so your knowledge base never gets contaminated with junk.

Embeddings

Cohere embed-multilingual-v3.0 (1024 dimensions) is the default embedder — covers multiple languages in the same vector store. VoyageAI is the configurable alternative.

Citations

Every RAG-based answer cites sources ([1] [2]), traceable back to the original chunk and source file/URL — that’s what separates “the agent answered” from “the agent answered based on something verifiable”.

Indexing

  • Drag a folder into the chat, or use /rag add.
  • The Memory (RAG) tab in the workbench shows what’s already indexed, lets you configure reranker on/off, top_k, embedding/rerank provider, and manage collections.
  • Indexing runs on an async queue (embedding_queue.db) — it doesn’t block chat while processing.

See also

Context Graph