Most AI finds the consensus. This one challenges it.
Agentic-Contrarian is a local, privacy-first multi-agent RAG pipeline that reads global news, detects editorial blind spots, and generates a structured contrarian briefing — with every claim traced back to a real source.
"It doesn't summarise the news. It audits it."
The v2.0 demo will show: clickable citations · evidence mapping with stance badges · PDF export · Dev mode toggle.
| Improvement | Detail |
|---|---|
| ✅ Clickable citations | Every [NEWS_X] in the report links to the real source URL |
| ✅ Hallucination guard | Auditor drops off-topic articles before the LLM ever sees them |
| ✅ Stance classification | Title + body analysed separately — detects mixed-stance articles |
| ✅ PDF export | Full briefing: report + evidence table + friction points, downloadable |
| ✅ Domain-agnostic pipeline | Works on any topic — no hard-coded domains or example queries |
| ✅ Evidence Mapping tab | Browse every retrieved article, filtered by stance, with source links |
| ✅ Developer mode | Debug tab hidden by default; enable with DEV_MODE=1 or ?dev=1 |
| ✅ Relevance gate | 40% keyword-match threshold + exact OpenSearch query (no fuzzy noise) |
Mainstream RAG systems are optimised for agreement — they retrieve the most popular view and amplify it. This creates three problems:
- Echo chambers — consensus sources dominate retrieval
- Hidden risk — dissenting signals get buried by relevance scoring
- Unverifiable claims — LLMs hallucinate sources users can't check
Agentic-Contrarian addresses all three: it actively seeks non-consensus sources, classifies stance before ranking, and cites every claim with a real URL.
User query
│
▼
┌─────────────────────────────────────────────────────┐
│ Researcher Node │
│ Expands query into 13 variants across 9 categories │
│ (mainstream, opinion, trade, analyst, skeptical…) │
│ Deduplicates by URL + title │
└─────────────────┬───────────────────────────────────┘
│ evidence pool (up to 40 articles)
▼
┌─────────────────────────────────────────────────────┐
│ Auditor Node │
│ • Relevance gate: drops articles < 40% keyword match│
│ • Stance classifier: supportive / skeptical / │
│ critical / mixed / neutral │
│ (title vs body disagreement → mixed) │
│ • Reranks by: relevance × stance diversity × │
│ source diversity + contrarian boost │
│ • Extracts friction points (consensus vs dissent) │
│ • Fills empty contrarian_evidence from pool │
└─────────────────┬───────────────────────────────────┘
│ top 12 ranked, classified articles
▼
┌─────────────────────────────────────────────────────┐
│ Perspective Analyst Node │
│ Identifies missing source types (trade press, │
│ regional outlets, analyst notes, interviews…) │
└─────────────────┬───────────────────────────────────┘
│ blind spot list
▼
┌─────────────────────────────────────────────────────┐
│ Contrarian Node (Llama 3.2) │
│ Generates 4-section briefing: │
│ Mainstream Narrative / Contrarian View / │
│ Evidence / Caveats │
│ Every sentence must cite a source_id │
│ Post-generation topic check guards off-topic output │
└─────────────────────────────────────────────────────┘
graph LR
User((User)) --> UI[Streamlit Dashboard]
subgraph Local_Infrastructure
UI --> API[FastAPI Backend]
API --> Graph[LangGraph Orchestrator]
Graph --> DB[(OpenSearch)]
Graph --> LLM[Ollama / Llama 3.2]
end
RSS[[RSS Feeds — BBC, NYT, Guardian,\nEconomist, TechCrunch, Brookings…]] --> DB
Every query returns a structured 4-section briefing:
## Mainstream Narrative
What most retrieved sources say about the topic [NEWS_1 · BBC].
## Contrarian View
Non-consensus interpretation from skeptical/critical sources [NEWS_3 · ForeignPolicy].
## Evidence
Specific facts supporting the contrarian view [NEWS_2 · TechCrunch].
## Caveats
Missing evidence, underrepresented stances, and reasons the contrarian view may be wrong.
Citations are clickable links in the dashboard. The full briefing exports as a structured PDF.
| Component | Technology |
|---|---|
| Multi-agent orchestration | LangGraph |
| LLM — local, zero-cost | Ollama — Llama 3.2 |
| Vector + keyword search | OpenSearch |
| API layer | FastAPI + Uvicorn |
| Frontend | Streamlit — dark mode |
| News ingestion | RSS feeds — no API key required |
| Containerisation | Docker |
| Tool | Required |
|---|---|
| Python 3.11 | ✅ |
| Docker Desktop | ✅ (OpenSearch + Postgres) |
| Ollama | ✅ |
git clone https://github.com/PreethaRaj/Agentic-Contrarian.git
cd Agentic-ContrarianThen run the setup script — it creates the venv, installs all dependencies, pulls the LLM, starts Docker, and ingests the first batch of news:
setup.bat
setup.batautomates:python -m venv venv→pip install -r requirements.txt→docker compose up -d→ollama pull llama3.2→python ingest.py
After setup, launch both the backend and dashboard in one step:
start.bat
start.batopens two terminal windows automatically:
- Terminal 1 — FastAPI backend on
http://localhost:8000- Terminal 2 — Streamlit dashboard on
http://localhost:8501
Verify the backend is alive: http://localhost:8000/health → {"status":"ok"}
Then open http://localhost:8501, enter any topic, click Launch Deep-Dive Investigation.
Re-run any time to pull the latest articles into OpenSearch:
venv\Scripts\activate
python ingest.pyset DEV_MODE=1
streamlit run app\ui\dashboard.pyOr append ?dev=1 to the dashboard URL without restarting.
Agentic-Contrarian works on any topic — the pipeline is domain-agnostic.
The best queries come directly from the news sources the system reads. Try this: open any of the supported RSS feeds below, pick a headline that interests you, and paste the topic into the dashboard. The system will surface whether there is a contrarian. Example : "Britain’s Electorate Is ‘Splintering.’ Can Its System Stand the Strain?"
view hiding beneath the mainstream coverage.
| Source | URL |
|---|---|
| BBC World | https://feeds.bbci.co.uk/news/world/rss.xml |
| The Guardian | https://www.theguardian.com/world/rss |
| New York Times | https://rss.nytimes.com/services/xml/rss/nyt/World.xml |
| TechCrunch | https://techcrunch.com/feed/ |
| The Economist | https://www.economist.com/the-world-this-week/rss.xml |
| Foreign Policy | https://foreignpolicy.com/feed/ |
| Brookings | https://www.brookings.edu/feed/ |
| Channel NewsAsia | https://www.channelnewsasia.com/rss-feeds/8395904 |
Tip: Queries with visible tension in the headline work best — e.g. "X is booming despite Y", "Country splits on Z", "Experts warn about W". These already signal that a contrarian view likely exists in the index.
| Category | Sources |
|---|---|
| Mainstream | BBC World, New York Times, The Guardian |
| Opinion | The Economist, Foreign Policy |
| Trade/Tech | TechCrunch |
| Analyst | Brookings Institution, CFR |
| Regional | Channel NewsAsia, Sky News |
To add sources: open ingest.py and add RSS URLs to the FEEDS dictionary.
Agentic-Contrarian/
├── app/
│ ├── agents/
│ │ ├── nodes/
│ │ │ ├── researcher.py # Query expansion (13 variants, 9 categories)
│ │ │ ├── auditor.py # Relevance gate + stance classifier + reranker
│ │ │ ├── perspective.py # Blind spot detection
│ │ │ └── contrarian.py # Briefing generation with citation enforcement
│ │ ├── graph.py # LangGraph wiring
│ │ └── state.py # AgentState TypedDict
│ ├── api/main.py # FastAPI endpoints
│ ├── config.py # All tuneable constants (thresholds, signals, variants)
│ └── ui/dashboard.py # Streamlit dashboard
├── ingest.py # Standalone RSS ingest — no Airflow required
├── tests/test_pipeline.py # 30 tests, 7 domains, no live services needed
├── docker-compose.yml
├── requirements.txt
└── README.md
All pipeline tuneables live in app/config.py — edit once, affects the whole pipeline:
MIN_RELEVANCE_SCORE = 0.40 # raise to reduce noise; lower to increase recall
TOP_K_RERANKED = 12 # articles passed to LLM
MIN_CONTRARIAN = 2 # minimum skeptical articles before weak-signal flag
SOURCE_CATEGORIES = [...] # query expansion variants
SKEPTICAL_SIGNALS = {...} # stance classification vocabulary- Inference latency: Llama 3.2 on CPU: 60–180s per query. GPU or cloud LLM reduces this dramatically.
- Index freshness: Index is static until you re-run
python ingest.py. No live feed polling. - Contrarian depth: Limited by what RSS feeds actually publish. Niche topics may return weak contrarian coverage.
- No auth or rate limiting: Not production-ready for public deployment as-is.
- Streaming SSE output — word-by-word dashboard updates
- RAGAS evaluation node — automated faithfulness + contrarian-ness scoring
- Scheduled ingest — cron/Airflow for automatic index refresh
- GPU-optimised inference config
- Additional feed categories: academic preprints, financial filings, government reports
New to agentic RAG? Follow this:
| Day | File | What to learn |
|---|---|---|
| 1 | ingest.py |
How RSS feeds are chunked, vectorised, and indexed |
| 2 | app/agents/nodes/researcher.py |
Query expansion — why 13 variants, not 1 |
| 3 | app/agents/nodes/auditor.py |
Relevance gating, stance classification, reranking |
| 4 | app/agents/nodes/contrarian.py |
Prompt engineering — how to enforce citation grounding |
| 5 | app/config.py |
Tune the pipeline — change thresholds, observe output differences |
Built locally. Runs privately. Cites everything.