Forensic narrative analysis for news intelligence.
Built for Web Data UNLOCKED — Bright Data's two-day enterprise AI hackathon in San Francisco, May 30–31, 2026.
Narrative Alpha ingests news articles about a target event from multiple outlets, extracts a structured knowledge graph from each, computes a consensus baseline, then surfaces exactly how each outlet deviated — factual omissions, linguistic spin, and outlier claims. It tracks outlet reputation over time, surfacing which sources consistently front-run what the consensus later acknowledges.
This is not a fact-checker or misinformation detector. It maps narrative topology — what the institutional press agrees on, who omitted what, and who spun it.
| What This Is NOT | What This Tries To Be | |
|---|---|---|
| Generic "AI research assistant" | Opinionated — makes explicit calls about omission, spin, and outlier provenance | |
| Scraping + summarization | Metricized — every output is a scored, measurable claim, not prose summary | |
| RAG over web data | Forensic rather than assistive — interrogates sources instead of answering user questions | |
| Autonomous browsing agents | Structurally decomposed — graph extraction and set math, not agentic tool-calling loops | |
| Market-news summarizers | Measures epistemic distortion across sources rather than retrieving relevant content | |
| Sentiment dashboards | ||
| Lead generation / SEO automation | ||
| Generic "multi-agent" wrappers |
Articles → Knowledge Graphs → Consensus Baseline → Distortion Matrix → Dashboard
- Ingest — Bright Data SERP API + Web Unlocker pull articles from multiple outlets
- Normalize — Fast LLM resolves entity synonyms across sources
- Extract — Reasoning LLM builds structured node-and-edge graphs per article
- Analyze — Set math identifies omissions, embedding distance catches spin, provenance tracking flags outliers
- Display — Static dashboard renders the forensic report
| Metric | What It Measures |
|---|---|
| Omission Index (Oᵢ) | What facts did this outlet leave out? |
| Framing Volatility (Vf) | How much spin did they layer on? |
| Scatter-Shot Anomaly (Sa) | Does this outlet pump out noise? |
| Consensus Baseline (Gc) | What does everyone agree on? |
| Layer | Technology |
|---|---|
| Scraping | Bright Data SERP API + Web Unlocker |
| LLM | Configurable per pipeline step — DeepSeek, OpenAI, Google, Groq (any OpenAI-compatible endpoint) |
| Embeddings | OpenAI text-embedding-3-small (hardcoded — see Known Limitations) |
| Backend | FastAPI + uvicorn (Python 3.11) |
| Storage | SQLite (WAL mode) |
| Frontend | React + Vite |
Finance & Market Intelligence — Web Data UNLOCKED Hackathon, 2026.
Requires: Python 3.11+, Node.js 18+, npm
git clone git@github.com:afshinator/narrative-alpha.git
cd narrative-alpha
cp .env.example .env
# Fill in BRIGHTDATA_API_KEY, DEEPSEEK_API_KEY, OPENAI_API_KEY
./start-demo.shThe script installs Python and dashboard dependencies on first run. Open the URL it prints when ready — backend on port 8000, dashboard on port 5173 by default.
Port overrides (set in .env or inline):
BACKEND_PORT=9000 PORT=3005 ./start-demo.shEmbeddings require OpenAI. Framing volatility scores are computed using embedding distance (text-embedding-3-small). This call is hardcoded to OpenAI — OPENAI_API_KEY is always required even if you configure every other pipeline step to use a different provider (DeepSeek, Groq, etc.). Making the embedding provider configurable is a planned improvement.
MIT