Anamnesa

Indonesia's national clinical guidelines live in long government PDFs published by the Ministry of Health. A primary-care clinician with a specific question often has one real option: open the PDF on a phone and search through hundreds of pages, sometimes with a patient still in the room. Commercial clinical-decision-support tools solve the interface, but they are English-only and do not cover Indonesian guidelines or the national formulary. I am an emergency physician and GP in Indonesia. Anamnesa is the retrieval layer that workflow is missing, built first for me, and for clinicians who work the way I do. Ask a question in Indonesian. An orchestrator routes it through four agents. A Normalizer (Haiku 4.5) extracts clinical intent from colloquial Indonesian and refuses on out-of-scope or patient-specific questions. A Retriever runs hybrid vector and lexical search against 9,083 chunks from 81 public-domain guidelines, with no language model involved at this stage. A Drafter (Opus 4.7, adaptive thinking, high effort) composes an Indonesian answer with inline citations, can narrow retrieval on its own if the initial chunks are insufficient, and never emits a claim without a source. A Verifier (Opus 4.7, 1M context) independently re-reads every cited chunk and classifies each claim as supported, partial, or unsupported. An unsupported claim gives the Drafter exactly one retry; if the retry also fails, the whole answer is refused rather than shipped. Every citation carries a currency flag, current, aging, or superseded by a newer edition, and clicking one opens the exact page of the source PDF inside the app.