Synthesis ⚗️

Turn videos, audio, and text into a structured, semantically searchable knowledge base - entirely on your own machine.

Pipeline: ingest → transcribe → distil → embed → search

Install · Quick start · How it works · Roadmap

Synthesis ingests a source, distils it into atomic insights, links those insights to what you already know, and lets you search across everything by keyword or meaning. No data leaves your machine.

Built for clinicians, clinical teams, public-sector analysts, and anyone navigating complex, information-dense problems in environments where sending data to third-party AI services is not an option.

Quick start

1. Install Elixir via Mise:

   curl https://mise.run | sh
   mise use --global elixir@latest erlang@latest

2. Install yt-dlp and Ollama:

   pip install yt-dlp
   ollama pull qwen3.6:27b qwen3-embedding:8b

3. Clone and run:

   git clone https://github.com/The-Strategy-Unit/synthesis.git
   cd synthesis
   mix deps.get
   mix wiki.add "https://www.youtube.com/watch?v=<id>"
   mix wiki.search "your query"

Refreshing links in an existing wiki

If you generated notes before the cross-presentation linking feature, or you want to refresh links after changing settings:

mix wiki.link         # recompute semantic links in the database
mix wiki.update_links # rewrite Related sections in all markdown notes

Publish your wiki with an interactive graph

Host your generated wiki as a static site with an Obsidian-like graph view using Quartz:

git clone https://github.com/jackyzha0/quartz.git
cd quartz
npm i
npx quartz create -s ./../output/
npx quartz build --serve  # preview locally at localhost:8080

Deploy the public/ output to Azure Static Web Apps (free tier) or GitHub Pages.

How it works

1. Ingest - paste a YouTube URL, audio file, or text.
2. Transcribe - extract a transcript locally.
3. Distil - a local model pulls out atomic insights and a summary.
4. Store - insights are saved with semantic embeddings in a local SQLite database.
5. Search - query by keyword or meaning across your whole knowledge base.

Privacy and data sovereignty

• All processing runs locally via Ollama.
• No API keys, no cloud dependency, no data egress.
• Supports approved enterprise providers such as Azure OpenAI in future releases.

Synthesis is not a medical device and must not process patient-identifiable data. It is a knowledge-management tool for educational and research use.

Install

Non-technical users

A one-click installer is on the roadmap. For now, follow the developer install above. We welcome feedback on what would make setup easier.

Developers

Synthesis is written in Elixir on the Erlang VM, which gives us lightweight concurrency, fault tolerance, and the ability to ship a standalone local-first binary.

Configuration is handled in config/config.exs or via .env.

Setting	Default	Notes
ollama_url	http://localhost:11434	Ollama host
ollama_model	qwen3.6:27b	Main inference model
ollama_model_embed	qwen3-embedding:8b	Embedding model
chunk_concurrency	2	Parallel chunks for long transcripts
single_chunk_threshold	2500 tokens	Below this, process in one call
max_retries	3	Retry attempts on LLM failure
output_dir	output/	Markdown notes location
db_path	synthesis.db	SQLite database location

Roadmap

• One-click installer for non-technical users
• Audio and plain-text ingestion
• Azure OpenAI support
• Shared team knowledge bases

License

See LICENSE.