MiniMed

MiniMed is an offline-first search and evidence assistant for medical source material. A clinician can paste a patient case, search a local SQLite corpus, open the exact source fragment, and optionally use a small local model to reorder results and extract source-linked diagnostic or dosing evidence.

The application remains useful with no network, model, account, or hosted backend.

Current boundary

Implemented:

• Russian case-query parsing, including negative findings and missing-field prompts;
• local SQLite/FTS5 and deterministic vector retrieval;
• exact document, section, chunk, anchor, and neighboring-context navigation;
• installable read-only content-module contracts;
• optional browser-local GGUF inference;
• model-assisted reranking and exact-source clinical extraction;
• deterministic rejection of invented citations and unsupported dose text.

The public pilot contains 15 source-linked navigation cards: seven clinical recommendations and eight official medication-registry identity records. These cards are enough to test retrieval, but they are not complete clinical sources. In particular, the installed pilot does not contain verified dosing regimens.

See PRODUCT.md for the product contract, CURRENT_STATE.md for implemented status, and TECHNICAL_PLAN.md for the ordered plan.

Quick start

Requirements:

• Bun 1.2.3;
• Python 3.12+ and uv for content tooling;
• Chromium for browser tests.

Install only locked dependencies:

bun install --frozen-lockfile
uv sync --project tools/ingest --all-groups --locked

Run the web app:

bun run dev

Local servers must bind to 127.0.0.1. Run app, test, and build processes from a sanitized environment that does not inherit provider credentials, release tokens, private-corpus paths, or upload targets. Local-model downloads are optional and disabled in browser automation unless explicitly enabled.

Verification

bun run check
bun run typecheck
bun run test
bun run build
bun run python:check
bun run benchmark:all
bun run native:source:check

Build and benchmark the public Russian pilot:

bun run content:lint:pilot
bun run content:build:pilot
bun run benchmark:pilot

The generated database and reports live under data/build/ and are intentionally ignored by Git. Never hand-edit generated SQLite or JSON packs.

Add private documents

Private inputs remain outside Git:

bun run content:prepare:private
bun run content:lint:private
bun run content:build:private

The deterministic preparer accepts declared Markdown, text-layer PDF, and OCR text inputs. It preserves checksums and page/block or line provenance before the pack builder creates SQLite.

Do not commit source documents, patient data, model weights, credentials, or generated private packs.

Repository map

apps/app/                 SolidJS browser application
apps/landing/             Static project page
packages/contracts/       Public typed contracts
packages/core/            MedicalCore orchestration and retrieval fusion
packages/search-lexical/  Russian query analysis
packages/search-semantic/ Deterministic portable vector profile
packages/storage*/        Store ports and SQLite/native adapters
content/                  Public authored and synthetic fixture sources
tools/ingest/             Deterministic content pipeline
tools/benchmarks/         Retrieval and clinical-case benchmarks
docs/                     Product, architecture, and operating contracts

Dependency direction is:

UI → MedicalCore → ports → adapters

private sources → deterministic preparer → Markdown/provenance → pack builder → SQLite