Robin — Hybrid Search & Ranking for Swiss Real Estate

Natural-language + image search over 25,546 Swiss rental listings, with explainable ranking in DE / FR / IT / EN.

📦 Install · 🚀 Usage · 🏗 Architecture · 🧪 Dev · ☁ Deploy · 🎯 Challenge

Given a query like "3-room bright apartment in Zurich under 2,800 CHF with balcony, close to public transport", Robin:

1. Extracts hard constraints with an LLM (rooms, price, city, required features).
2. Gates retrieval through SQL — a listing that violates a hard constraint is structurally unable to reach the result set.
3. Fuses four retrieval channels inside the gate — BM25 (FTS5), dense text embeddings, SigLIP-2 text→image, DINOv2 image→image — via Reciprocal Rank Fusion (k=60).
4. Blends 8 normalized signals into a final score with templated, explainable reasons.
5. Remembers each authenticated user's saves / likes / clicks to personalize the next query (opt-in).

Design doc: ARCHITECTURE.md. Built for the ETH/Uni Datathon 2026 in partnership with robinreal.

• Hard-filter gate — constraint violations are impossible, not merely unlikely
• 4-language query understanding — DE / FR / IT / EN, LLM-structured outputs
• Hybrid retrieval — BM25 + Arctic-Embed-L v2 + SigLIP-2 Giant + DINOv2 ViT-L/14
• Image-to-image search — upload any photo, find listings that look like it
• Commute-aware ranking — ~125 k real transit-minute pairs (r5py over SBB GTFS)
• Enrichment — 166-column side table, 100% coverage with explicit UNKNOWN sentinels (no fabrication)
• Session-scoped personalization — save / like / dismiss → memory-based re-ranking
• Explainable — every ranked listing carries a human-readable reason

Every result shows what we understood (LLM-parsed hard constraints), hard-filter pass/fail, the BM25 keywords found in the listing text, the real travel-minutes to user-picked landmarks (ETH, UZH, Altstadt, Zürich HB) from the GTFS commute graph, and the four fused channels with their contribution to the final score.

A horizontal bar chart exposes each signal's contribution to the final score (Words · Photos · Meaning · Your wishes). Hard-filter pass/fail and visual / semantic match scores sit in the footer. Love / Save / Bring back feed the personalization memory.

Every transit minute Robin shows is auditable — one click opens the Google Maps route from the listing address to the selected landmark. Users can cross-check our GTFS-derived r5py graph against Google's routing, so the commute signal isn't a black box.

flowchart LR
    Q["/ user query + optional photo /"] --> LLM["LLM query understanding"]
    LLM --> GATE{{"SQL hard-filter gate"}}
    GATE --> BM25["FTS5 BM25"]
    GATE --> DENSE["Arctic-Embed-L v2"]
    GATE --> SIG["SigLIP-2 Giant"]
    GATE --> DINO["DINOv2 ViT-L/14"]
    BM25 --> RRF["Reciprocal Rank Fusion (k=60)"]
    DENSE --> RRF
    SIG --> RRF
    DINO --> RRF
    RRF --> BLEND["8-signal linear blend<br/>price · transit · POI · commute · …"]
    MEM[("user memory<br/>saves · likes · dismisses")] -.opt-in.-> BLEND
    BLEND --> OUT["top-K listings + reasons"]

    classDef gate fill:#fee,stroke:#c00,color:#900
    classDef ml fill:#eef,stroke:#44c,color:#228
    class GATE gate
    class LLM,DENSE,SIG,DINO,RRF,BLEND ml

See ARCHITECTURE.md for the full design, evidence, and references.

git clone https://github.com/<org>/Datathon_2026
cd Datathon_2026

cp .env.example .env            # fill in your API keys
uv sync --group image_search    # ~1 GB torch + transformers

# Get the pre-built dataset bundle (~2.5 GB, see docs/DATASET.md):
uv run --with gdown gdown 1rxYd17eX7C99AgNVlJft1QIFGtwogxX8 -O datathon2026_dataset.zip
unzip datathon2026_dataset.zip  # into the repo root

uv run uvicorn app.main:app --reload
open http://localhost:8000/demo

Full setup walk-through: docs/INSTALLATION.md.

Python 3.12 · FastAPI · SQLite (FTS5) · Snowflake Arctic-Embed-L v2 · Google SigLIP-2 Giant · Meta DINOv2 ViT-L/14 · r5py + SBB GTFS · OpenAI Structured Outputs · MCP Apps SDK · React + Vite (widget)

Module reports (cross-check against the design doc):

• robinreal — challenge sponsor and listing corpus
• Comparis & SRED — additional listing sources
• SBB / OpenTransportData.swiss — GTFS feed (CC-BY)
• OpenStreetMap — Nominatim + Switzerland extract (ODbL)
• Snowflake — Arctic-Embed-L v2 (Apache 2.0)
• Google Research — SigLIP-2 (Apache 2.0)
• Meta FAIR — DINOv2 (Apache 2.0)

MIT — see LICENSE.