Introduction
Vedana is an open-source framework for building AI assistants whose answers can be verified and reproduced. Vedana stores domain knowledge as a knowledge graph, and the assistant explores this graph step by step through explicitly described tools — vector search and Cypher queries against the graph database.
Unlike classic RAG, Vedana:
- doesn’t guess the answer based on top-K similar text fragments — it builds a structured query;
- returns complete sets (not samples), exact values (not approximations), and explicit sources (not “somewhere in the documents”);
- lets you describe your domain as a data model (anchors, attributes, links, queries) and get deterministic behaviour.
Who needs this
Vedana is designed for domains where mistakes are not allowed or are expensive:
- e-commerce — product catalogs, compatibility, in-store availability, delivery policies;
- legal and compliance — which requirements apply to a product, which documents regulate a category;
- internal knowledge bases — answers based on documentation, regulations, organisational structure;
- customer support — stable answers to FAQs, intent-based routing, exact values instead of “roughly that”.
If your assistant’s answers must be verifiable and aligned with real business logic — Vedana is your case. If “summarize this PDF” is enough, classic RAG will do.
What’s in the box
- JIMS — a framework for managing threads, events, and pipelines. Supports several interfaces: Telegram, Terminal UI, HTTP API, web widget, Reflex backoffice.
- Vedana Core — a RAG pipeline with data model filtering, an agent, and the
cypherandvector_text_searchtools. - Vedana ETL — an incremental pipeline on top of Datapipe for loading the data model and data from Grist into Memgraph and pgvector.
- Vedana Backoffice — a Reflex admin UI with a chat, an ETL pipeline inspector, an evaluation harness, and prompt settings.
- LiteLLM support: you can use OpenAI, OpenRouter, Google/VertexAI, and any compatible providers.
- Observability: OpenTelemetry traces, Prometheus metrics, Sentry integration.
What’s next
- Quick Start — bring Vedana up locally and ask your first question.
- Concepts — the theory: what Semantic RAG is, why classic RAG breaks, how the data model is organised.
- Architecture Overview — how the code is structured and how the components fit together.