A self-hosted knowledge workspace that turns team content into source-grounded, permission-aware knowledge for people and AI agents.
Akasha helps organizations turn scattered work context and experience into knowledge that can be discovered, connected, reused, and grounded in its sources. Teams collaborate in shared spaces, while agents retrieve and work with organizational knowledge within the same permission boundaries.
It brings together a collaborative Wiki, AI-powered knowledge compilation and retrieval, relationship-aware navigation, and MCP-based agent access in a self-hosted platform.
- Table of Contents
- Why Akasha
- Features
- How It Works
- Benchmark Results
- Roadmap / Vision
- Development
- Agent Integration
- Self-hosted
- Acknowledgements
- Contributors
Most knowledge tools store pages, while most AI assistants retrieve isolated fragments. Akasha connects source material, derived knowledge, people, and agents in one permission-aware system.
- Wiki-first organizational memory — Teams keep a familiar collaborative workspace while building a reusable knowledge layer above it.
- Evidence before answers — Derived knowledge and AI responses remain connected to source pages and supporting evidence.
- Structure beyond similarity — Links, entities, claims, concepts, and relations provide context that keyword or vector similarity alone cannot capture.
- One knowledge surface for people and agents — Human and MCP access follow the same workspace and resource permissions.
- Infrastructure you control — Akasha is self-hosted, with configurable storage and model endpoints.
| Area | Capabilities |
|---|---|
| Collaborative workspace | Spaces, rich-text and Markdown pages, attachments, comments, version history, real-time collaboration, and access control |
| Knowledge compilation | Asynchronous page and space compilation into entities, concepts, claims, relations, comparisons, contradictions, and evidence |
| Retrieval and Q&A | Lexical and vector retrieval, source citations, supporting evidence, and permission-aware results |
| Relationship graph | Visual exploration of direct page links and compiled semantic relationships |
| Agent integration | MCP access to knowledge and permitted page, space, comment, attachment, and workspace operations |
| Self-hosting | PostgreSQL with pgvector, configurable model endpoints, and local, S3, or Azure file storage |
Akasha keeps source content and derived knowledge as distinct but connected layers.
1. Capture trusted sources. Teams create Wiki pages or import content into shared spaces. Original pages, attachments, and their permissions remain the canonical source layer.
2. Compile structure and evidence. Akasha asynchronously turns selected pages and spaces into entities, concepts, claims, relations, comparisons, and contradictions. Each artifact retains source references and supporting evidence, while indexes make the compiled layer searchable. Compilation augments the original Wiki rather than replacing it.
3. Retrieve and act. People search and ask questions through the workspace; agents use MCP. Lexical, vector, and relationship-aware retrieval assemble relevant context, while source permissions continue to govern results. Answers can cite supporting pages, and supported agent operations are recorded for auditing.
Akasha is evaluated on three public multi-hop question-answering datasets using the qwen-27b model. R@5 measures the proportion of gold evidence retrieved in the top five results, while F1 balances the precision and recall between answer and reference answer.
| Dataset | R@5 | F1 |
|---|---|---|
| MuSiQue | 89.33 | 67.01 |
| 2WikiMultiHopQA | 91.00 | 71.95 |
| HotpotQA | 99.00 | 87.82 |
| Average | 93.11 | 75.59 |
Under the benchmark configuration currently recorded in this repository, Akasha achieved an average Recall@5 (R@5) of 93.11 across MuSiQue, 2WikiMultiHopQA, and HotpotQA, with an average F1 of 75.59. These figures describe the recorded retrieval run and should be compared only under consistent dataset and evaluation settings.
See the complete benchmark comparison and local evaluation toolkit for baselines, per-dataset results, and instructions for running the evaluation pipeline.
The following ideas describe Akasha's product direction. They are exploratory plans and should not be read as a list of features guaranteed to be available in the current release.
Akasha aims to evolve from a knowledge workspace into a broader organizational memory system:
- Factual memory — what happened, supported by source artifacts and provenance;
- Interaction memory — why decisions, disagreements, and trade-offs mattered;
- Action memory — what actions, workflows, and safeguards should follow.
Akasha may eventually allow reusable agent skills, operating patterns, and execution feedback to accumulate across tasks and agents, so that organizational experience becomes easier to reuse over time.
git clone https://github.com/chaterm/Akasha.git
cd Akasha
pnpm installThis repository is a pnpm workspace monorepo. Use pnpm for dependency installation and scripts.
Create the local environment file:
cp .env.example .envGenerate a local application secret and set it as APP_SECRET in .env:
openssl rand -hex 32Do not commit .env or any production credentials to the repository.
Start PostgreSQL with pgvector using the included Compose service, and provide Redis separately:
docker compose up -d db
docker run -d --name akasha-redis -p 6379:6379 redis:7Then apply migrations and start the development servers:
pnpm --filter ./apps/server run migration:latest
pnpm run devOpen http://localhost:3000. For environment requirements, model configuration, service details, and troubleshooting, see docs/development.md.
pnpm run build # Build all workspace projects
pnpm run client:build # Build the frontend
pnpm run server:build # Build the backendBuild artifacts are generated under the corresponding apps/*/dist and packages/*/dist directories.
Akasha provides an MCP endpoint for agents to access the knowledge workspace.
Configure the MCP server with:
- the absolute URL of the deployed Akasha instance followed by
/mcp; - an API key with the required workspace permissions.
The MCP integration supports knowledge queries and permitted operations on pages, spaces, comments, attachments, and workspace information. Requests follow Akasha's authorization rules.
See akasha-plugin/README.md for installation instructions and host-specific configuration examples.
Akasha is designed to run in self-hosted environments. Organizations can control where application data, file storage, and AI model endpoints are configured, while applying their own access-control and operational policies.
Akasha builds upon excellent open-source projects. We gratefully acknowledge:
- Docmost — the collaborative Wiki foundation that the workspace and editor layers build on.
- OpenDataLoader PDF — the PDF parsing engine behind document import.
Thank you for your contribution! Please refer to the Contribution Guide for more information.