build(deps-dev): bump eslint-plugin-react-hooks from 5.2.0 to 7.1.1 in /web/dashboard

.github/workflows/check.yml

@@ -20,6 +20,7 @@ jobs:
       - run: go test -short -race -count=1 ./cmd/... ./internal/...
       - run: go vet ./cmd/... ./internal/...
       - run: go build ./cmd/api
+      - run: make wasm-check
 
   dashboard:
     runs-on: ubuntu-latest
@@ -38,6 +39,13 @@ jobs:
       - run: npm run lint
       - run: npm test
       - run: npm run build
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: Build the in-browser runtime (WASM core)
+        working-directory: .
+        run: make wasm-build
       - run: npx playwright install --with-deps chromium
       - run: npm run e2e
 

.gitignore

@@ -76,3 +76,7 @@ __pycache__/
 
 # Local dynamic-int8 reranker export produced by sidecars/reranker/quantize.py
 int8_avx2/
+
+# Built by `make wasm-build` for the in-browser runtime
+web/dashboard/public/nadir/
+web/dashboard/.wrangler/

AGENTS.md

@@ -71,6 +71,33 @@ The React dashboard remains in `web/dashboard`, and the Python reranker and
 optional Docling processes remain in `sidecars/`. Compose builds `cmd/api`, so
 normal local and Compose startup use the same backend.
 
+## Browser runtime (WASM)
+
+`cmd/wasm` builds the same graph for the browser (ADR 0038): `make wasm-build`
+writes `.local/wasm/` and `web/dashboard/public/nadir/`; `make wasm-check`
+builds, vets and tests it under wasm plus a Node smoke test of the whole API.
+Select it in the dashboard with `?runtime=browser` (`&models=mock` for scripted
+models, `&llm=none` for retrieval only) or `VITE_RUNTIME=browser`. Keep
+`internal/bootstrap/assembly` and `internal/core/**` free of gRPC, OS and
+server-only imports so they stay buildable for js/wasm.
+
+Testing the browser runtime (`make` targets from the repo root, `npm`/`npx` from
+`web/dashboard`):
+
+| Layer | Command | Needs |
+|-------|---------|-------|
+| Go + wasm + Node API smoke test | `make wasm-check` | Go, node |
+| Store benchmark (exact vs HNSW) | `go test -run '^$' -bench HybridSearch -benchtime=100x ./internal/providers/memory/documents` | about 5 min; builds up to 50k chunks |
+| Dashboard unit tests | `npm test` | node |
+| Browser e2e, scripted models | `make wasm-build`, then `npx playwright test e2e/browser-runtime.spec.ts` | Chrome; no downloads or GPU |
+| Real embedder + reranker | `E2E_REAL_MODELS=1 npx playwright test e2e/browser-runtime-real.spec.ts` | ~330 MB download, cached in `.local/pw-real-models-profile` |
+| Real WebLLM answer | `E2E_REAL_LLM=1 npx playwright test e2e/browser-runtime-llm.spec.ts` | WebGPU and a ~1.2 GB download; set `E2E_LLM_MODEL` to change the model |
+
+Try it by hand: `make wasm-build`, then `npm run dev` and open
+`/?runtime=browser` (add `&models=mock` to skip downloads). Headless Chrome on
+Linux needs the Vulkan flags the LLM spec sets to see a GPU, and drivers without
+WebGPU `shader-f16` need a `q4f32_1` model id (default builds are `q4f16_1`).
+
 ## Key rules
 
 - Core use cases must not import `internal/edge/http/`, `internal/providers/`,

Makefile

@@ -1,4 +1,4 @@
-.PHONY: run compose compose-rerank test race vet build benchmark benchmark-ui eval eval-generate help check mdn generate-mocks
+.PHONY: run compose compose-rerank test race vet build wasm-build wasm-smoke wasm-check benchmark benchmark-ui eval eval-generate help check mdn generate-mocks
 
 GO_PACKAGES := ./cmd/... ./internal/...
 MOCKERY_VERSION ?= v2.53.7
@@ -26,6 +26,28 @@ vet:
 build:
 	go build ./cmd/api
 
+WASM_DIR := .local/wasm
+
+# Browser runtime: cmd/wasm assembles the retrieval and indexing graph over
+# in-memory storage and host-supplied model functions.
+wasm-build:
+	mkdir -p $(WASM_DIR)
+	GOOS=js GOARCH=wasm go build -o $(WASM_DIR)/nadir.wasm ./cmd/wasm
+	cp "$$(go env GOROOT)/lib/wasm/wasm_exec.js" $(WASM_DIR)/wasm_exec.js
+	mkdir -p web/dashboard/public/nadir
+	cp $(WASM_DIR)/nadir.wasm $(WASM_DIR)/wasm_exec.js web/dashboard/public/nadir/
+
+# End-to-end Node smoke test of the built wasm (ingest, query, snapshot, reset).
+wasm-smoke: wasm-build
+	node cmd/wasm/smoke.mjs $(WASM_DIR)/wasm_exec.js $(WASM_DIR)/nadir.wasm
+
+# Keeps the provider-independent graph buildable and runnable as WASM (needs
+# node). The assembly test drives ingest, chunking, embedding, storage and
+# search through in-memory providers inside the wasm runtime.
+wasm-check: wasm-smoke
+	GOOS=js GOARCH=wasm go vet ./internal/core/... ./internal/bootstrap/assembly ./internal/providers/browser ./internal/providers/memory/... ./cmd/wasm
+	GOOS=js GOARCH=wasm go test -count=1 -exec="$$(go env GOROOT)/lib/wasm/go_js_wasm_exec" ./internal/bootstrap/assembly ./internal/providers/memory/...
+
 benchmark:
 	$(BENCHMARK_PYTHON) -m benchmark $(ARGS)