Storming Bastille maps causal chains between historical events — not just facts, but the relationships between them. Five AI agents coordinate to produce a streamed narrative, an interactive causal graph, a zoomable timeline, and verified source citations.
You ask "What led to the French Revolution?" and within seconds Storming Bastille starts building a structured causal map: a streamed narrative backed by live web research, an interactive graph of connected events with confidence-weighted edges, a zoomable D3 timeline, and source citations ranked by reliability. Five AI agents coordinate behind the scenes, each doing one job well. The result is a session you can explore, edit, name, and revisit — not a paragraph you copy and paste.
Built around a single idea: causal relationships between events matter more than chronological facts.
React Flow canvas with five relationship types — direct cause, contributing factor, enabling condition, consequence, feedback loop — each with a confidence score. Drag, edit, and export to PNG.
Narrative text streams token by token via SSE. Graph nodes appear within 2 seconds. Causal edges arrive as a second pass. You watch the analysis build itself — never stare at a spinner.
A dedicated SourceVerifier agent cross-checks claims and assigns reliability tiers: Academic, Encyclopedia, News, Web, Blog. Citations appear in a ranked panel alongside the graph.
Events plotted chronologically on a D3 SVG timeline. Zoomable, pannable, and linked to the graph — click a timeline event to highlight its corresponding node.
Name and save any analysis. Sessions persist server-side with SQLite and automatic GCS backup. Pick up where you left off, or run follow-up queries that build on prior context.
Five specialists: HistorianAgent (Haiku), CausalAnalystAgent (Sonnet), GraphBuilderAgent, SourceVerifierAgent, and FollowUpAgent. Each optimised for its precise task.
Designed to minimise time-to-first-pixel. Something meaningful is on screen within 2 seconds of submitting a query.
DuckDuckGo and Bing are queried simultaneously. Results are injected as plain text into the HistorianAgent prompt — not via tool-use, to avoid latency overhead. HistorianAgent (Claude Haiku 4.5) streams the narrative and embeds 8–10 structured events as JSON.
GraphBuilderAgent converts the historian's structured events into React Flow nodes and Dagre-positioned timeline data. An SSE event fires — nodes appear on screen before any causal analysis begins. No AI call needed; this is pure deterministic transformation.
CausalAnalystAgent (Claude Sonnet) identifies relationships between events: type, confidence score (0.3–1.0), and explanation. A second SSE event fires with the edge set. The graph completes visually as edges appear between the existing nodes.
SourceVerifierAgent cross-checks claims and assigns reliability tiers. Results are saved to SQLite with a GCS backup. This runs in the background — the user has everything they need before this phase completes.
A carefully chosen stack with no unnecessary dependencies.
Prerequisites: Python 3.12+, uv, and Bun.
Pull the source code and enter the project directory.
Copy the backend env file and set your Anthropic API key. That's the only required configuration to start.
The start.sh script boots both backend and frontend. Default login: admin / changeme.
# 1. Clone $ git clone https://github.com/rupakc/storming-bastille.git $ cd storming-bastille # 2. Configure $ cp backend/.env.example backend/.env $ echo "ANTHROPIC_API_KEY=sk-ant-..." >> backend/.env # 3. Run $ ./start.sh # Open → http://localhost:3000 (login: admin / changeme)
Full setup instructions including GCP deployment in the Setup and Installation wiki page.
Clone the repo, set your API key, and start exploring any historical question in seconds.
Everything you need to understand, extend, and deploy the project.