SynSearch

iGEM 2026 Software — Open-source RAG system for searching the iGEM knowledge archive.

A wet-lab biologist types a plain-English question. SynSearch retrieves the most relevant past iGEM team wikis and other relevant sources and returns a sourced answer with direct links to the original pages.

---

What's in this repository

app.py                  # Main Streamlit app (all pages)
requirements.txt        # Python dependencies
.streamlit/config.toml  # Dark theme config
harvest_corpus.py       # Converts Munich's xlsx → corpus.json (run once)
build_index.py          # Chunks + embeds corpus.json → Chroma index (run once, add later)
retrieval.py            # Search + similar-projects functions (add later)

---

Run locally

pip install streamlit
streamlit run app.py

No API key needed for the demo. The full pipeline additionally requires:

pip install sentence-transformers chromadb pandas requests beautifulsoup4

---

Building the real index

1. Download processed_data.xlsx from the Munich 2024 repo (see Attribution below)
2. Run python harvest_corpus.py → produces corpus.json
3. Run python build_index.py → produces a local Chroma vector store
4. Uncomment the import in app.py and remove the demo data block

---

What we built vs. what we inherited

Inherited (Munich 2024, CC BY 4.0):

• Pre-scraped 2019 iGEM wiki corpus (343 teams, processed_data.xlsx)
• Original scraper logic (process_data.py)

Built by us:

• Inline source citations with direct wiki links (Munich listed this as future work)
• Metadata filtering by year, track, and medal before retrieval
• "Similar projects" semantic explorer — not available anywhere else
• Three-way RAGAS benchmark: bare GPT-4o vs GPT-4o+RAG vs open model+RAG
• Zero-install hosted interface — no Docker, no user API key required
• Lighter stack: sentence-transformers + Chroma replaces Docker + Qdrant + Groq

---

Attribution

The 2019 iGEM wiki corpus used in this project was scraped and curated by the Munich 2024 iGEM team and is used under Creative Commons Attribution 4.0 International (CC BY 4.0).

Source: https://gitlab.igem.org/2024/software-tools/munich

---

Benchmark

System	Faithfulness	Answer Relevancy	Context Recall
Bare GPT-4o	0.41	0.58	0.29
GPT-4o + SynSearch RAG	0.79	0.83	0.71
Open model + SynSearch RAG	0.74	0.77	0.68

Evaluated on 50 manually-verified iGEM-specific questions using RAGAS. (placeholder — replace with real numbers when evaluation is run)

---

License

MIT. The Munich 2024 corpus retains its original CC BY 4.0 licence.