Zero Data Egress
MyKB runs entirely within your VPC or on your local machine. Keep proprietary code and docs off third‑party clouds.
The On-Prem AI Knowledge Engine
MyKB turns your private docs and code into a secure, auditable, self-healing intelligence engine. No data egress. No vendor lock-in. Your knowledge, on your metal.
Initial OSS release coming September 2025.
Deterministic & Auditable
Commit‑pinned citations and lineage give repeatable answers with proof — built for compliance and critical workflows.
No Runaway Costs
Deploy on fixed‑cost infrastructure you already own. Avoid per‑token surprise billing and lock in predictable spend.
One Engine, Tailored Solutions
Core Engine Features
Hybrid Search (Dense + Sparse)
Combines semantic understanding (FastEmbed/ONNX) with keyword precision (BM25) for superior relevance, fused with Reciprocal Rank Fusion (RRF).
Incremental Ledger (SQLite)
Content hashes detect chunk-level changes, ensuring only deltas are re-indexed. This makes ingestion efficient and idempotent.
Structure‑Aware Chunking
A Markdown AST parser creates stable, meaningful chunks that respect headings, code blocks, and tables for better context retrieval.
Zero‑Trust Auth Core
Production-grade identity using EdDSA (Ed25519) JWTs, revocable refresh tokens, device fingerprinting, and per-route rate limiting.
Manual & Automated Self‑Healing
OSS users can add synonyms/boosts via a `patches.json` file. Enterprise adds a UI-driven feedback loop for continuous, automated improvement.
Portable & Performant
Built with Python, FastAPI, and Qdrant. Runs efficiently on CPU with optional GPU acceleration. Deploys anywhere with Docker.
Unlock Your AI Workforce
MyKB's core engine powers specialized Co-Pilots—thin, secure agents that automate complex workflows.
Analyst Co-Pilot
Automates market research by ingesting external data and comparing it against internal roadmaps, securely.
Tools Used:
kb.seed.preview, kb.ingest, kb.search
Developer Co-Pilot
Accelerates development by providing instant, cited answers from the entire codebase and internal documentation.
Tools Used:
kb.search_code, kb.get, kb.sources
Compliance Co-Pilot
Enforces governance by checking documents and code against policies, using curated filters and self-healing patches.
Tools Used:
kb.search, kb.admin.patch, kb.explain_policy
One Core, Two Paths
| Feature | OSS (Solo & Team) | Enterprise |
|---|---|---|
| Core Engine | ||
| Full RAG Pipeline | ✓ | ✓ |
| Hybrid Search & Reranking | ✓ | ✓ |
| Incremental Ingestion Ledger | ✓ | ✓ |
| Security & IAM | ||
| JWT Auth (RBAC) | Coming Soon | ✓ |
| Zero-Trust Gateway | ✗ | ✓ |
| SSO/SAML/OIDC Integration | ✗ | ✓ |
| Automation & Ops | ||
| Manual Healing (patches.json) | ✗ | ✓ |
| Automated Self-Healing Loop | ✗ | ✓ |
| SLAs & Dedicated Support | ✗ | ✓ |
Frequently Asked Questions
Does my data ever leave my infrastructure?
Never. MyKB is architected to be on-premise first. The entire RAG pipeline—from ingestion to retrieval—runs on your hardware. You can optionally connect to an external LLM API for answer synthesis, but the retrieved context is the only data sent, and you can also use fully local LLMs for a 100% air-gapped deployment.
How does MyKB handle updates to my documents?
Our incremental ledger uses content hashing (e.g., git commit hashes or file checksums) to track changes at the document and chunk level. When you re-run the ingestion process, only new or modified chunks are processed and embedded, making updates extremely fast and efficient.
What vector database and models do you use?
We use Qdrant as our vector store due to its performance, on-disk storage capabilities, and support for named vectors. For embeddings, we default to high-performance, small-footprint models via FastEmbed (ONNX runtime), which are ideal for on-prem CPU deployments. However, the architecture is modular, allowing you to plug in other models.
What does "self-healing" actually mean?
Self-healing refers to our system for improving search relevance without retraining models. When a search fails or returns a poor result, an admin can create a "patch." In the OSS version, this is a rule in a JSON file (e.g., mapping a synonym like "k8s" to "kubernetes" or boosting a specific document). The Enterprise version provides a UI for this, and uses user feedback to automatically suggest and apply these patches.
Get Started with the Future of Private AI.
Join our community of privacy-first teams building the future of on-premise AI.