Vedana Documentation
Licensed under the Apache License 2.0 — free for commercial and non-commercial use, including modification and redistribution.
Vedana is an open-source multi-agent system for building AI assistants on top of Semantic RAG (Retrieval-Augmented Generation over a knowledge graph). Unlike classic RAG, Vedana doesn’t guess answers based on text similarity — it explores the knowledge base step by step: it issues Cypher queries against the graph, runs vector search, verifies sources, and assembles an answer from real data.
Vedana is built around three components:
- JIMS (Just an Integrated Multiagent System) — a framework for managing message flows, events, threads, and LLM providers.
- Vedana Core — the RAG pipeline with data model filtering, Cypher generation, and vector search.
- Vedana ETL — an incremental pipeline built with Datapipe that syncs the data model and the data itself from Grist into Memgraph and pgvector.
Where to start
Depending on your role:
- I’m a developer and want to run Vedana locally — start with Quick Start to see Vedana working in Docker. If you then want to develop on Vedana itself (debug a service, add a feature), see Local Development, then Architecture Overview.
- I’m a product manager and want to understand what this product is — start with What is Vedana and Why Classic RAG Fails, then Use Cases.
- I want to describe my domain and load data — read Data Model and then Data Ingestion.
- I want to contribute — see the Contributing Guide.
Documentation map
Getting Started
- Introduction — what Vedana is in one paragraph and who needs it.
- Quick Start — bring up the stack in Docker and ask your first question in 10 minutes.
- Local Development — develop on Vedana itself: native Python with
uv, infrastructure in Docker. (For first-run demo see Quick Start; for production see Operations → Deployment.) - Configuration — every key ENV variable and what it means.
Concepts (for all roles)
- What is Vedana
- Why Classic RAG Fails
- Semantic RAG Overview
- Data for Vedana
- Data Model for Vedana
- Tools for Vedana
- Playbook for Vedana
Architecture (for developers)
- Overview — high-level diagram, repository, component relationships.
- JIMS Core — threads, events, pipelines, ThreadController and ThreadContext.
- Vedana Core —
RagPipeline,RagAgent,LLM,Graph,VectorStore. - Vedana ETL — Datapipe catalog, steps, incremental loading.
- Vedana Backoffice — Reflex admin UI, chat, ETL runner, metrics.
- Storage Model — Postgres, Memgraph, pgvector, Grist.
- Observability — OpenTelemetry, Prometheus, Sentry.
Data Model
- Overview
- Anchors (nodes)
- Attributes
- Links (relationships)
- Queries (playbook)
- Prompts
- ConversationLifecycle
Data Ingestion
API Reference
- Overview
- HTTP API (
jims-api) - Widget API (
jims-widget) - Telegram (
jims-telegram) - Python API (
vedana-core) - Configuration Reference (
.env)
Guides
- Setting Up Data Model
- Test Dataset (LIMIT)
- Adding Anchors
- Adding Attributes
- Adding Links
- Adding Documents
- Adding Structured Data
- Adding FAQ Entries
- Tuning Embeddings & Thresholds
- Customizing Prompts
- Writing a Custom Tool
- Multi-tenancy
Product (for PMs)
- Use Cases
- Comparison with Classic RAG
- Quality Metrics & Evaluation
- Limitations
- FAQ
- Open roadmap items on GitHub
Operations
Contributing
License and community
The project is distributed under the Apache License 2.0 — a permissive license that allows commercial use, modification, and redistribution, including in proprietary products, with attribution and a notice of changes. Source code and issue tracker — at github.com/epoch8/vedana. Product website: vedana.tech.