🧠 Synthora — AI Research Agent

A fully self-contained AI research agent in a single Streamlit file. No FastAPI. No separate server. No local GPU. One command to run.

Submit a research query → get ranked sources, a structured summary, fact-checked claims, and a full downloadable markdown report — powered by the Ollama Cloud API (gpt-oss:120b-cloud).

---

🖼️ Output Preview

---

📊 Sample Output

---

🚀 Features

Feature	Detail
☁️ Ollama Cloud API	Direct HTTP integration — gpt-oss:120b-cloud or any configured model
🔍 Web search	DuckDuckGo scraping with automatic redirect URL decoding
📄 Content extraction	Full page text via BeautifulSoup (noise-stripped)
📊 LLM source ranking	Relevance-scored (0–1) and sorted
🧠 Structured summary	Key insights, statistics, arguments, risks, opportunities
✅ Fact-checking	Per-claim LLM verification with confidence score + status badge
📄 Report generation	Full markdown report: executive summary, findings, analysis, citations
⬇️ Report download	One-click .md download button
🎛️ Source count control	Sidebar slider — 3 to 10 sources
💾 JSON persistence	Flat file storage under data/ — no database required
📚 Research history	All past reports loadable from the history panel
⚡ Single file	Entire app in streamlit_app.py

---

🏗️ Architecture

Everything runs in one Python process. The Ollama Cloud API is called directly — no intermediate server layer.

streamlit_app.py
│
├── ollama_generate()       — POST to Ollama Cloud /generate
├── extract_json()          — robust JSON parser for LLM output
├── test_api_connection()   — verify API key + connectivity
│
├── search_agent()
│   ├── _ddg_search()       — DuckDuckGo HTML scrape
│   ├── _decode_ddg_url()   — decode ?uddg= redirect to real URL
│   ├── _extract_page()     — fetch + BeautifulSoup content extract
│   └── LLM ranking call    — relevance_score per source
│
├── summarize_agent()       — LLM → keyInsights/statistics/arguments/risks/opportunities
├── fact_check_agent()      — LLM per-claim: status + confidence + explanation
├── report_agent()          — LLM → full markdown report with citations
│
├── run_pipeline()          — orchestrates all stages + live progress bar
│
├── save_project()          — data/projects/project_*.json
├── save_report()           — data/reports/report_*.json
└── load_history()          — reads all project files, sorted by created_at

---

🛠️ Tech Stack

Layer	Technology
UI + App	Streamlit ≥ 1.35
LLM	Ollama Cloud API (gpt-oss:120b-cloud)
Web search	DuckDuckGo HTML (no API key needed)
HTML parsing	BeautifulSoup4
HTTP client	requests
Storage	JSON files (data/projects/, data/reports/)
Language	Python 3.11

Removed from previous version: FastAPI, Uvicorn, Pydantic, python-multipart, openai SDK, asyncio, separate backend process.

---

⚙️ Installation

git clone https://github.com/QubitGaurav/AI_Research_Agent.git
cd AI_Research_Agent

python -m venv venv
# Windows
venv\Scripts\activate
# Linux / macOS
source venv/bin/activate

pip install -r requirements.txt

---

🔑 Configuration

Copy the example and fill in your Ollama Cloud credentials:

cp .env.example .env

.env:

OLLAMA_API_KEY=your_ollama_api_key_here
OLLAMA_BASE_URL=https://ollama.com/api
OLLAMA_MODEL=gpt-oss:120b-cloud
OLLAMA_TIMEOUT=300
DATA_DIR=data

Variable	Default	Description
OLLAMA_API_KEY	—	Required. Your Ollama Cloud API key
OLLAMA_BASE_URL	https://ollama.com/api	Ollama Cloud base URL
OLLAMA_MODEL	gpt-oss:120b-cloud	Model for all LLM calls
OLLAMA_TIMEOUT	300	Request timeout in seconds
DATA_DIR	data	Where project/report JSON files are stored

---

▶️ Run

streamlit run streamlit_app.py

Open http://localhost:8501.

That's it. One process, one terminal.

---

🚀 Deploy to Streamlit Cloud

1. Push repo to GitHub (.env and data/ are in .gitignore — they won't be pushed).
2. Go to share.streamlit.io → New app → connect your repo.
3. Under App settings → Secrets, paste:

OLLAMA_API_KEY  = "your_ollama_api_key_here"
OLLAMA_BASE_URL = "https://ollama.com/api"
OLLAMA_MODEL    = "gpt-oss:120b-cloud"
OLLAMA_TIMEOUT  = "300"

No backend deployment needed. Streamlit Cloud talks directly to Ollama Cloud.

---

📁 Project Structure

.
├── streamlit_app.py          # entire application
├── requirements.txt          # 4 dependencies
├── runtime.txt               # python-3.11
├── .env.example              # copy to .env and fill in
├── .gitignore
│
└── .streamlit/
    └── secrets.toml          # Streamlit Cloud secrets (do not commit)

data/                         # auto-created at runtime
├── projects/                 # project_<id>.json — full pipeline state
└── reports/                  # report_<id>.json  — generated reports

---

🔄 Pipeline Stages

#	Stage	What happens
1	Search	DuckDuckGo HTML scrape → decode ?uddg= redirect → fetch real page → extract body text
2	Rank	LLM scores each source for relevance (0–1); sorted descending
3	Summarize	LLM extracts key insights, statistics, arguments, risks, opportunities as JSON
4	Fact-check	LLM verifies each claim against source excerpts; returns status + confidence
5	Report	LLM writes full markdown report: executive summary → findings → analysis → citations
6	Persist	Project + report saved to data/ as JSON; visible in history panel

---

⚠️ Known Limitations

Limitation	Notes
DuckDuckGo rate limiting	If you get CAPTCHA or empty results, wait a few minutes. For production volume, use a paid search API (Serper, SerpAPI).
Single model for all stages	All calls use OLLAMA_MODEL. To use different models per stage, add a model parameter to ollama_generate().
No streaming	Ollama /generate is called with stream: false — the UI updates only after each full stage completes.
File storage not concurrent-safe	Fine for personal/single-user use. For multi-user production, replace save_project/save_report with SQLite.
Fact-check quality	Per-claim verification is only as good as the source content extracted. Paywalled pages return empty content.

---

👨‍💻 Author

Gaurav Sharma GitHub: QubitGaurav