detroit.dev

A community knowledge base that remembers everything.

detroit.dev is an open, AI-powered knowledge base built and maintained by the detroit.dev community. Members contribute markdown notes on any topic — research summaries, meeting notes, podcast transcriptions, tutorials — and the system makes all of it searchable, queryable, and accessible from the web, Discord, or the command line.

Ask a question in plain English. Get an answer grounded in what the community actually wrote.

---

Why this exists

Knowledge dies in chat scroll. Good conversations happen in Discord, someone explains something well in a meeting, a podcast covers exactly the right topic — and then it's gone. detroit.dev captures that knowledge into a permanent, searchable, AI-queryable corpus that the whole community can build on.

What you can do with it

Search	Semantic search across the entire corpus. "How do we keep fuel available?" matches tritium-breeding content — no exact keywords needed.
Ask	Chat with the corpus. The AI retrieves the 5 most relevant chunks and synthesizes an answer grounded in community knowledge.
Capture	Transcribe podcasts, record meeting notes, save Discord threads — all become searchable corpus entries.
Read	Every note renders as a clean, readable article with adjustable fonts, themes, and a table of contents.
Contribute	Add a markdown file, open a PR, and your knowledge is part of the shared brain.

Quick start

git clone https://github.com/JacobHind/detroitdotdev.git
cd detroitdotdev
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Build the search index

# Requires Ollama running locally (https://ollama.com)
ollama pull nomic-embed-text
python tools/ingest.py

Search and chat

python tools/search.py "how do reactors breed fuel"
python tools/serve.py          # opens http://localhost:8888 with AI chat sidebar

Transcribe audio

python tools/transcribe.py recording.mp3 --qa    # local, no API key needed

Build the static site

python tools/build_site.py     # renders corpus → site/

Architecture

┌─────────────────────────────────────────────────────────────┐
│                        Interfaces                           │
│  Web reader (/site)  │  Discord bot (/bot)  │  CLI tools   │
└──────────┬───────────┴──────────┬───────────┴──────┬───────┘
           │                      │                  │
           ▼                      ▼                  ▼
┌─────────────────────────────────────────────────────────────┐
│                     RAG Engine (common.py)                   │
│  Chunk docs → Embed → Index → Retrieve → LLM answer         │
└──────────┬──────────────────────────────────────────────────┘
           │
           ▼
┌─────────────────────────────────────────────────────────────┐
│                      Corpus (/corpus)                        │
│  Markdown files organized by topic, contributed by members   │
└─────────────────────────────────────────────────────────────┘

How search and chat work

1. Ingest — ingest.py splits every markdown file into semantic chunks (~800 words each, split at headers), embeds them with nomic-embed-text, and saves the vectors to a JSON index.
2. Search — search.py embeds your query with the same model and ranks chunks by cosine similarity.
3. Chat — serve.py / api/chat.py / Discord bot retrieve the top 5 chunks for your question, pass them to an LLM as context, and return a grounded answer.

Project structure

corpus/                  Community-contributed markdown notes
  topics/                Organized by topic (ai-ml, nuclear-fusion, web-dev, etc.)
  members/               Community member pages
  transcripts/           Podcast/video transcriptions
  generated/             Embedding index (auto-generated, gitignored)

tools/                   The engine
  common.py              Chunking, embedding, vector search, RAG retrieval
  ingest.py              Build the semantic search index
  search.py              CLI semantic search
  serve.py               Dev server with RAG chat at /api/chat
  transcribe.py          Audio/video → markdown + Q&A (faster-whisper, local)
  build_site.py          Render markdown → readable HTML articles
  chat.py                Standalone chat client

bot/                     Discord bot (see bot/README.md)
  bot.py                 /ask, /search, /capture, /notes commands

api/chat.py              Vercel serverless endpoint (production)
templates/               HTML templates for the static site
site/                    Generated static site output

Infrastructure

Component	How it runs	Cost
Embeddings	nomic-embed-text via Ollama (local)	Free
Vector index	JSON file + numpy cosine similarity	Free
LLM	Ollama (local), or Groq / OpenRouter / OpenAI	Free or BYOK
Transcription	faster-whisper on CPU (local)	Free
Discord bot	Any server / VPS / Railway	Free–$5/mo
Website	Vercel (static + serverless)	Free tier

Zero mandatory API keys. Everything runs locally with Ollama. Optionally set GROQ_API_KEY, OPENAI_API_KEY, or OPENROUTER_API_KEY for cloud LLMs.

Discord bot

The bot brings the entire knowledge base into your Discord server:

Command	What it does
/ask [question]	RAG-powered Q&A — retrieves relevant chunks, sends to LLM, answers in-channel
/search [query]	Semantic search — top 5 results with relevance scores and previews
/capture [title]	Save recent channel messages as a corpus note (creates a GitHub PR)
/notes	List recent corpus additions

Setup: see bot/README.md.

Contributing

See CONTRIBUTING.md for the full guide. The short version:

1. Add a .md file under corpus/topics/<category>/
2. Include frontmatter (title, author, date, tags)
3. Run python tools/ingest.py to re-index
4. Open a PR

Moderators: CONTRIBUTING.md has a dedicated section on content review, quality standards, and how to manage the corpus.

License

• Content (corpus/) — CC BY-SA 4.0 (share and adapt with attribution)
• Code (tools/, bot/, api/) — MIT