KB - Personal Knowledge Base

A local-first, AI-powered knowledge base that understands your documents, conversations, and projects.

Overview

KB (Knowledge Base) is a personal, single-user system that ingests your documents and chat history, extracts meaningful information, and builds a semantic understanding of your knowledge.

Instead of manually organizing notes, KB continuously learns from your data - so future interactions are smarter, more contextual, and personalized.

Key Features

• AI-Powered Knowledge Extraction.
  Automatically analyzes documents and conversations to extract relevant insights.
• Document Ingestion Pipeline.
  Upload raw files - KB processes, chunks, and indexes them for semantic search.
• Chat Memory Integration.
  Conversations are summarized and merged into your evolving knowledge base.
• Semantic Search (Vector-Based).
  Uses embeddings to retrieve relevant context instead of keyword matching.
• Local-First Architecture.
  Your data stays on your machine. No external vector DB required.
• Context-Aware Prompting.
  Retrieved knowledge is injected into LLM prompts to improve response quality.

Architecture

 ┌──────────────────────┐
│ Frontend │
│ (In Development) │
└─────────┬────────────┘
│
▼
┌──────────────────────┐
│ .NET Backend │
└─────────┬────────────┘
│
┌─────────────────┼─────────────────┐
▼ ▼ ▼
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ Ingestion │ │ Chat Flow │ │ Retrieval │
│ Pipeline │ │ │ │ │
└──────┬───────┘ └──────┬───────┘ └──────┬───────┘
▼ ▼ ▼
┌──────────────────────────────────────────────────┐
│ SQLite Vector Store │
│ - Embeddings │
│ - Summaries │
│ - Extracted user knowledge │
└──────────────────────────────────────────────────┘

How It Works

1. Ingestion
   Documents are uploaded
   Content is chunked (typically 512 tokens)
   Embeddings are generated and stored

2. Knowledge Extraction
   LLM analyzes content and extracts:

   • facts
   • key points
   • decisions
   • user-specific insights

3. Chat Processing
   Conversations are:

   • summarized
   • merged into long-term memory
   • indexed for retrieval

4. Retrieval + Prompting
   Relevant context is fetched via semantic search
   Injected into system prompt
   LLM generates context-aware responses

Tech Stack

Backend: .NET 10
AI Integration: Microsoft.Extensions.AI
Vector Store: SQLite (local-first)
Frontend: In progress

Getting Started

Warning

Project is in active development

Prerequisites

• .NET 10 SDK
• API key for your LLM provider (if applicable)

Setup

git clone https://github.com/sys27/kb.git
cd kb/backend
dotnet build

Run

dotnet run

Design Principles

• Local-first — your data belongs to you
• Single-user focused — optimized for personal knowledge, not teams
• Composable architecture — easy to swap components (LLM, embeddings, storage)
• Incremental intelligence — the system improves as you use it

Roadmap

• Frontend UI
• Advanced retrieval (hybrid search: vector + keyword)
• Better knowledge extraction pipelines
• Better ingestion system
• Migration path to external vector DB (e.g., Qdrant)

License

GNU GPL v3. See the LICENSE file for details.