// the find
truefoundry/cognita
RAG (Retrieval Augmented Generation) Framework for building modular, open source applications for production by TrueFoundry
Cognita is a Python framework that wraps LangChain/LlamaIndex to give RAG pipelines a production-ready structure: modular parsers, loaders, embedders, and query controllers all exposed via FastAPI, with a React UI for non-technical users. It targets teams trying to graduate from Jupyter notebooks to something deployable. The project is explicitly abandoned — the README leads with a deprecation notice.
The separation between the indexing job and the query server is correct architecture — they have genuinely different scaling profiles and failure modes. Multiple vector DB backends (Qdrant, Milvus, SingleStore, Weaviate, Mongo) are actually implemented, not just listed. Incremental indexing that diffs the vector DB state against the data source before re-embedding is a real time-saver that most RAG tutorials skip entirely. The decorator-based query controller pattern (`@query_controller`, `@post`) is a clean way to let teams add custom retrieval logic without touching the framework internals.
The project is dead — maintainers posted a deprecation notice and the last meaningful activity was mid-2024. Do not build on this. Using Prisma (a Node.js ORM) as the metadata store in a Python service is a genuinely strange choice that adds a Node.js dependency and a separate migration workflow to a Python stack. The whole thing sits on top of LangChain and LlamaIndex, so you're inheriting two notoriously unstable abstraction layers with their own breaking-change cadences. The TrueFoundry deployment path is first-class and the self-hosted path is clearly secondary — the docs assume you'll eventually pay them.