← Back to the Suite
Knowledge & Analytics

Your codebase.
Recalled.

Full RAG stack — ingestion, chunking, embedding, vector search — scoped to your project and organized with tags. Your agents ask, AgentCognos answers. Retrieval as a solved problem.

Get Started Free →
AgentCognos — MCP Interface
# Ingest project documentation
recall_ingest(
  source: "./docs/architecture.md",
  project: "agentquanta-backend",
  tags: ["architecture", "backend", "infra"]
)
 
✓ Ingested 847 chunks · 3 tags · project: agentquanta-backend
 
# Retrieve with project + tag scoping
recall_search(
  query: "how does the websocket auth handshake work",
  project: "agentquanta-backend",
  tags: ["architecture"],
  limit: 5
)
 
✓ 5 results · ranked by relevance · scoped to project
 
  [1] docs/architecture.md:§4.2 · score 0.94
     architecturebackend
     "The QuantaHub WebSocket upgrade validates the cert pin before
      accepting the connection. Clients present their leaf cert..."
 
  [2] docs/security.md:§2.1 · score 0.87
     architectureinfra

Know what you built.
Instantly.

AgentCognos handles the entire knowledge pipeline — from raw files to ranked, tagged, project-scoped retrieval — behind a single MCP interface your agents can call like any other tool.

📂
Project-Scoped Context

Every document, chunk, and embedding is namespaced to a project. Agents working on agentquanta-backend only search that project's knowledge base — no cross-contamination, no irrelevant results from other repos or contexts.

🏷️
Tag-Based Organization

Every chunk carries the tags assigned at ingest time. Search can filter to a specific tag or combination — ["auth", "backend"] narrows results to exactly the right slice of your knowledge base. Tags are first-class, not metadata afterthoughts.

⚙️
Full Ingest Pipeline

Point AgentCognos at a file, directory, or URL. It handles reading, chunking (with configurable overlap and chunk size), embedding via Ollama (nomic-embed-text, 768-dim), and storage to pgvector. The entire pipeline runs locally — your data never leaves your machine.

🔍
Hybrid Vector + Full-Text Search

Results are ranked by combining pgvector cosine similarity with PostgreSQL full-text search using Reciprocal Rank Fusion. Semantic understanding catches concept matches even when keywords don't align. Exact keyword matching catches what embeddings sometimes miss.

🔌
Single MCP Interface

Every operation — ingest, search, list projects, list tags, delete, re-embed — is available as an MCP tool. Your agents call recall_search the same way they call any other tool. No SDK, no HTTP client, no credentials to manage in the agent prompt.

🔄
Live Re-ingest

Files change. Re-ingesting a source automatically supersedes stale chunks — old embeddings are marked, not deleted, so you can audit what changed. Version-aware chunking tracks modification timestamps, making it safe to re-run ingest on an evolving codebase at any time.

Local RAG.
No cloud required.

AgentCognos runs entirely on your machine. PostgreSQL + pgvector for storage, Ollama for embeddings. If Ollama is unavailable, search falls back to full-text automatically.

Storage & Retrieval
  • PostgreSQL + pgvector, HNSW index, cosine distance
  • 768-dim embeddings via Ollama (nomic-embed-text)
  • Hybrid search: cosine + full-text RRF scoring
  • Project namespacing — results scoped per project
  • Tag filtering — first-class, not metadata
  • Configurable chunk size and overlap
  • Stale chunks marked on re-ingest, not deleted
MCP Tools
  • recall_ingest — file, directory, or URL
  • recall_search — query + project + tags + limit
  • recall_list_projects — all indexed projects
  • recall_list_tags — tags within a project
  • recall_delete — remove by source or project
  • recall_status — index health and chunk counts
  • Works with Claude Code, any MCP-compatible agent

Part of the AgentQuanta Suite

AgentCognos pairs naturally with AgentSynapse — Synapse for episodic agent memory, Recall for structured knowledge retrieval. Use both and your agents always have the right context.