Cymatix-Context: Deterministic RAG for Agents with SQLite

Overview

Cymatix Context is a local-first retrieval system for agents, built on SQLite FTS5 BM25. It indexes code and documents in a SQLite knowledge store, then retrieves and assembles context without an LLM call on the retrieval path. Its know / miss response contract is intended to help agents distinguish grounded context from a retrieval miss rather than guess.

It provides CLI, HTTP, and MCP interfaces over the same retrieval primitives. The project is aimed at developers and agent builders who want a self-hosted context source, with deterministic retrieval behavior and an explicit signal when the store does not support an answer.

Key Features

• Ingests code and documents into a local SQLite knowledge store.
• Uses FTS5 BM25, tags, synonym expansion, and ranking stages for retrieval; dense BGE-M3 recall is optional and off by default.
• Returns a structured know or miss result, with freshness checks that can turn stale results into misses.
• Offers CLI, HTTP, and MCP access to retrieval and packet-building functions.
• Compresses and assembles retrieved context to fit a configured token budget, with an experimental opt-in full-text companion profile.
• Includes session delivery that can elide documents already delivered in a session.

Use Cases

• Give coding agents a searchable, local index of a repository when answers should be grounded in project files.
• Provide Claude Code, Codex, Gemini CLI, or other MCP clients with context from a shared local knowledge store.
• Add retrieval to scripts or CI workflows through the CLI, without running a model for each query.
• Serve context to an IDE or an agent using the HTTP API, including an OpenAI-compatible proxy workflow.

Getting Started

Python 3.11 or later is required. For the recommended CPU and MCP working set:

pip install "cymatix-context[cpu,mcp]"
python -m spacy download en_core_web_sm

Then ingest a project and query it:

cymatix ingest path/to/your/project/ --recursive
cymatix query "how does this project work?"

See the README and setup guide for configuration and integration details.

Alternatives

• ragtime: Ragtime offers a self-hosted RAG API and chat UI with FAISS or pgvector indexes and model integrations, rather than SQLite FTS5 retrieval without an LLM call.
• cq: cq focuses on coding agents sharing and validating reusable lessons, rather than general-purpose local knowledge retrieval through CLI, HTTP, and MCP.
• ReMe: ReMe manages searchable, editable Markdown memory from agent conversations and resources, rather than retrieving knowledge with SQLite FTS5 BM25.
• utopia: Utopia organizes knowledge as an ontology-guided, bitemporal graph with evidence-backed answers and human review, rather than a local BM25 search index.

Considerations

• The project is young, although its measured activity indicates regular commits and releases and ongoing issue and pull-request handling. Its contributor activity is highly concentrated in one person, which is a maintenance and continuity risk.
• SQLite is the local store. The core install is described as dependency-light; optional embeddings pull in PyTorch, and the README says dense recall is off by default.
• The README reports a substantial accuracy gap for the sharded path and recommends unsharded use for accuracy-sensitive corpora. Several benchmark categories also remain weak.
• Cymatics reordering is explicitly experimental, and the confidence value is not yet presented as a reliable trust signal.
• The project is licensed under Apache-2.0. Review the license and third-party attributions in NOTICE when incorporating it.