Multi Search Server

Search Google, Bing, and DuckDuckGo at the same time in real Chrome windows, then merge the organic results into a clean, ad-free consensus view — with a local Ollama model summarizing the overlap and picking a best match.

The consensus pane is a small FastAPI app. You type once there; Playwright drives typed searches in the other three windows (so you stay in a normal browser session / profile instead of hitting raw search APIs).

┌─────────────┬─────────────┐
│   Google    │    Bing     │
├─────────────┼─────────────┤
│ DuckDuckGo  │  Consensus  │  ← FastAPI UI + Ollama summary
└─────────────┴─────────────┘

Why this exists

Comparing engines by hand means juggling tabs and ads. Multi Search Server:

1. Opens four visible Chrome windows (2×2 on a big screen)
2. Types your query into three engines like a person would
3. Pulls organic results (and follows Next / More results once for more depth)
4. Merges by URL / title overlap and ranks consensus hits
5. Asks local Ollama (gemma4:e4b by default) for a short summary + best match
6. Shows everything in a sticky-search-bar consensus UI (summary scrolls with results)

Requirements

• Python 3.11+ (3.12 recommended)
• Google Chrome (or Chromium)
• A desktop session (headed Chrome — not pure SSH without a display)
• Ollama running locally, with a model pulled (default: gemma4:e4b)
• Optional: Rectangle (macOS) or wmctrl / xdotool (Linux) for auto 2×2 tiling

Install

git clone https://github.com/deneskelenyi/Multi-Search-Server.git
cd Multi-Search-Server

python3.12 -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e .

System Chrome is used via Playwright’s channel=chrome — you do not need playwright install for Chromium unless you prefer it.

Usage

# Start Chrome + FastAPI consensus UI at http://127.0.0.1:8765
python -m search_consensus

# Optional initial query (auto-runs when the UI loads)
python -m search_consensus "best espresso machine under 200"

# Custom bind / Ollama
python -m search_consensus --host 127.0.0.1 --port 8765 \
  --ollama-model gemma4:e4b --ollama-url http://127.0.0.1:11434

# Attach to Chrome already started with remote debugging
#   chrome --remote-debugging-port=9222 --user-data-dir="$HOME/.search-consensus-chrome"
python -m search_consensus --cdp http://127.0.0.1:9222

# Use a real Chrome profile (close Chrome first)
python -m search_consensus --profile Default

# Skip automatic window tiling
python -m search_consensus --no-tile

Type in the Consensus window search box. That POSTs to /api/search, which searches the other three windows, merges results, and refreshes the summary + list.

HTTP API

Method	Path	Purpose
GET	/	Consensus UI (?q= optional initial query)
GET	/health	Liveness / engine readiness
POST	/api/search	Body: {"query": "..."} → consensus JSON + Ollama summary

Chrome profiles

Mode	Behavior
(default)	Dedicated profile under ~/.search-consensus-chrome
--profile Default	System Chrome Default profile (close Chrome first)
--cdp URL	Attach to an already-running Chrome with remote debugging

Using your real profile helps avoid bot walls; engines can still show CAPTCHAs — leave the windows visible so you can solve them.

Project layout

search_consensus/
  app.py            # Orchestrates Chrome windows + uvicorn
  server.py         # FastAPI routes
  engines.py        # Typed search + parsers + pagination
  consensus.py      # Merge / rank logic
  ollama_client.py  # Local LLM summary
  ui.py             # Consensus HTML/CSS/JS
  chrome.py         # Persistent profile / CDP launch
  tiling.py         # macOS / Linux 2×2 layout helpers
  cli.py            # `python -m search_consensus`

Notes

• Google Split View is only two panes; this project uses four windows instead.
• Page-size URL hints (num=, count=) are best-effort; clicking Next once is the reliable way to deepen results.
• Works on macOS and Linux with the same Python code.

License

Use and modify as you like for personal or experimental tooling. Respect each search engine’s terms of service.