knowledge-base

Personal Knowledge Base

A minimal note-taking web app with an AI chat that can search across your notes. Notes live in a Supermemory bash sandbox under /notes/, so the same sgrep / cat / ls tooling powering the agent's search is what stores your data — no separate database, no embeddings setup.

Features

• 📝 Add, browse, view, and delete markdown notes from the sidebar.
• 💬 Chat with Claude, which has a bash tool wired to your notes directory and can semantically search them with sgrep.
• 📡 Streaming responses over Server-Sent Events with live tool-call blocks so you can see what the assistant is doing.
• 🪶 Pure HTML / CSS / vanilla JS frontend — no build step.

Prerequisites

• Python 3.10+
• A Supermemory API key
• An Anthropic API key

Setup

cd examples/knowledge-base

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate

pip install -r requirements.txt

cp .env.example .env
# then edit .env and fill in your two API keys

Usage

python server.py

Then open http://localhost:8000 in your browser.

1. Click + Add Note in the sidebar to create a few notes (markdown is fine — the title becomes /notes/<title>.md).
2. Type a question in the chat. The assistant will call the bash tool, run sgrep / cat against /notes/, and stream back an answer that cites the source note(s).

Architecture

┌─────────────────────────┐         ┌─────────────────────────┐
│  Browser (vanilla JS)   │ <──SSE─ │  FastAPI server         │
│  - notes sidebar        │ ──HTTP─►│  - /api/notes CRUD      │
│  - chat + tool blocks   │         │  - /api/chat (stream)   │
└─────────────────────────┘         └────────────┬────────────┘
                                                 │
                                                 ▼
                                  ┌──────────────────────────┐
                                  │ supermemory_bash sandbox │
                                  │  /notes/*.md             │
                                  │  + sgrep / cat / ls      │
                                  └──────────────────────────┘
                                                 ▲
                                                 │ tool_use
                                  ┌──────────────┴───────────┐
                                  │ Anthropic Claude         │
                                  │ (agent loop, ≤10 steps)  │
                                  └──────────────────────────┘

• server.py exposes the REST + SSE API and serves static/.
• Each request creates a fresh bash instance scoped to the knowledge_base container tag, so all notes live in a single persistent sandbox keyed by that tag.
• The chat endpoint streams Claude's text and tool-use events as SSE so the UI can render them progressively.

Tech stack

• FastAPI + Uvicorn
• supermemory-bash
• Anthropic Python SDK (Claude Sonnet 4)
• Vanilla HTML / CSS / JavaScript (no build step)

Notes

• Slashes in note titles are replaced with _ so the path stays a single file under /notes/.
• The agent loop is capped at 10 iterations per chat turn.
• Conversation history is kept client-side and sent with every request; refreshing the page starts a new conversation but your notes persist in the sandbox.