MAPR — tactical OSINT command console

Live global-event intelligence: a Rust worker ingests real news feeds, enriches (geocode / severity / NER / correlation) and embeds them, writing into a self-hosted Convex backend that serves a lean React console with a bespoke MapLibre GL 3D map and a grounded RAG assistant — all self-hostable from one docker compose.

Layout (monorepo)

Dir	What
convex/	Serving/RAG/auth/realtime backend — self-hosted Convex (TS functions). Schema, queries/mutations/actions, Stripe http.ts, crons, Convex Auth.
ingestor/	Rust worker: fetch (GDELT/RSS/HTML, SSRF-safe) → dedup → geocode → severity → NER → correlate → embed (Ollama bge-m3) → idempotent ingestBatch.
web/	React 19 + Vite + convex/react. MapLibre GL map (flat + globe), bottom-center composer (deterministic filter + grounded RAG), region/trends/entities/intel/workspace/admin.
migration/	One-shot legacy → Convex backfill (re-embed) + Stripe relink. Reads data/.
deploy/	One-command self-host bundle: docker compose up -d → Convex + Ollama + ingestor + web + tunnel + bootstrap. See deploy/README.md.
design-ref/	Design + contract reference (tokens, parity checklist, Convex contract, Phase-0 verification).
data/	Legacy SQLite — the migration's historical-backfill source (only needed for the one-shot cutover; otherwise the live ingestor populates everything).

Quick start (production, single box)

cp deploy/.env.example deploy/.env   # fill secrets (see deploy/README.md)
docker compose -f deploy/docker-compose.yml --env-file deploy/.env up -d

Ollama auto-pulls bge-m3 + qwen2.5:3b, the bootstrap deploys the Convex functions + seeds real sources, and the ingestor begins populating live data.

Local development

• Backend: cd convex && bun install && bunx convex dev (against a local self-hosted Convex backend — see deploy/PHASE0-VERIFICATION.md).
• Web: cd web && bun install && bun run dev.
• Ingestor: cd ingestor && cargo run -- --once (needs CONVEX_URL, OLLAMA_URL, MAPR_INGEST_KEY).

AI / data flow

Rust ingestor → ingestBatch (Convex) → reactive feed. The composer parses intents deterministically (drives the map, no LLM) or routes free-form questions to a grounded, citation-enforced RAG action (bge-m3 retrieval + qwen2.5 generation via Ollama). One Ollama instance serves both embeddings and generation.