Search the bridge,
not the web.
A search interface over every document AMB has indexed — ministry strategic plans, surveillance bulletins, programme reviews — plus every published dossier and session note. Default mode returns a citation list of exact passages. Optional AI synthesis returns a one-paragraph answer with every claim citable to a chunk.
Citation-list default
Every search returns the exact passages from the indexed corpus, with document name, page and a stable permalink. No model interpretation unless you ask for it.
AI synthesis · opt-in
Toggle on synthesis and Claude Haiku writes a one-paragraph answer using only the retrieved chunks as context. Every claim is footnoted to a chunk you can verify.
Cross-country search
Ask “EPI coverage drop-out 2021–2024 across pilot countries” and see the relevant Tanzania, Malawi and DRC chunks side by side.
Bilingual EN/FR
French queries find French documents. Embeddings are language-aware. DRC documents and English documents share the same retrieval surface.
Public-domain only
M13 indexes only what M1 publishes. Take-down requests are honoured within 5 business days; the embedding index is rebuilt nightly.
API access
M20 exposes M13 results as JSON for partner directories (MAP, IDEM-LAB, Vector Atlas, AIMS alumni). Anyone can ingest AMB into their own tooling.
At build time, every Layer-0 document and every dossier is chunked, embedded with Cloudflare's BGE-base-en model on Workers AI, and stored in Vectorize. The query worker accepts natural-language queries, embeds them, retrieves top-k chunks. The default response is a deterministic citation list — no model interpretation. AI synthesis is a separate, opt-in Claude Haiku call that always shows the underlying chunks. Default-to-citation respects the openness pledge and avoids hallucination liability.
The live module in v0.0 is M10 · Document Triage.
A working demo of the same back-end architecture. Try it.