Vault
Your knowledge, amplified.
Vault is the knowledge retrieval system powered by RAG. Ingest your documents, and Pulse searches them with hybrid vector + full-text search to ground its responses in your actual data.
Documents are ingested through an intelligent pipeline: source files are parsed, split into optimal chunks, embedded into high-dimensional vectors, and stored in PostgreSQL with pgvector.
Search uses a hybrid approach combining dense vector similarity (semantic meaning) with BM25 sparse matching (exact keywords). This ensures relevant results whether the query uses the same terminology or describes the concept differently.
Knowledge Pipeline
Document Ingestion
Upload text, PDF, Markdown, and code files. The ingestion pipeline extracts content and prepares it for embedding.
Smart Chunking
Documents are split into optimal chunks with overlap. Section boundaries, paragraphs, and code blocks are respected.
Multi-Provider Embedding
Choose from OpenAI, Together, Voyage, Cohere, or local embedding models. Different dimension/quality tradeoffs.
pgvector Storage
Vectors stored in PostgreSQL with HNSW index for fast approximate nearest neighbor search. No separate vector database needed.
Hybrid Search
Dense vector similarity + BM25 sparse matching. Find relevant chunks by semantic meaning, exact keywords, or both.
Proactive Context Injection
Relevant document chunks are automatically injected into the system prompt for every request — alongside Engram memories. RAG happens transparently.
Incremental Updates
Add, update, or remove documents at any time. The index updates incrementally — no full rebuild required.
Provider Flexibility
Switch embedding providers without re-ingesting documents. Vault supports migration between providers with re-embedding tools.
Multi-Format Support
Plain text, Markdown, PDF, HTML, and source code files. Each format has a specialized parser for optimal content extraction.
How Vault processes knowledge
Document uploaded
You upload a file or paste text. The ingestion pipeline detects the format and extracts clean content.
Chunked and embedded
Content is split into overlapping chunks and each chunk is embedded into a high-dimensional vector.
Stored in pgvector
Chunks and vectors are stored in PostgreSQL with HNSW indexing for fast approximate nearest neighbor retrieval.
Retrieved at query time
When you ask a question, Vault searches your documents with hybrid search and injects the most relevant chunks into the AI context.