Summary

When a user uploads a file via the UI, today they see ~4 plaintext messages over the whole indexing job and silent failures on LLM 5xx during extraction. This PR replaces the upload SSE channel with structured per-phase events.

• Pipeline (internal/pipeline/pipeline.go) — extends ProgressEvent with File, ChunksDone, ChunksTotal. indexFile now emits at every phase boundary: load → chunk → embed (with chunk counters) → extract_entities → extract_relationships → extract_claims → structure. Send is non-blocking so a slow consumer never stalls indexing. Existing CLI callers pass nil Progress and remain unaffected.
• Upload handler (internal/api/handlers.go) — per-job structured event log (jobProgressLog w/ sync.Cond). The upload goroutine creates a buffered chan pipeline.ProgressEvent, passes it to IndexOptions, and a relay goroutine fans events into the log. The SSE handler at GET /api/upload/progress?job_id=... now writes data: {json}\n\n frames where {json} is {job_id, file, phase, chunks_done, chunks_total, message, done, error}. Indexing failures emit a terminal phase: "error" event. Legacy plain-text relay preserved when ?job_id= is omitted. Route path unchanged.
• UI hook (ui/src/hooks/api/useUploadProgress.ts) — fetch + ReadableStream-based SSE consumer with a shared store accessed via useSyncExternalStore. Robust frame parser (parseSseChunk) handles split frames, comment/retry/id lines, and falls back gracefully on plain-text legacy frames.
• Modal (ui/src/routes/documents/UploadModal.tsx) — per-file row per phase (Linear-style restraint, semantic shadcn classes, aria-live="polite"), embed N/M chunks counter, inline text-destructive for failures.

End-to-end verified against the agent's ollama/bge-m3 + llama3.2 setup: 11 distinct phases observed for a single-file upload, terminating on done:true. Production server on :37777 was not touched.

Test plan

• make build clean
• make vet clean
• make test — all packages green
• cd ui && npm run build — TypeScript + Vite clean
• cd ui && npm test -- --run — 102/102 tests pass (incl. new useUploadProgress.test.tsx)
• Existing TestUploadProgress_JobIDFiltering regression test extended; new TestUploadProgress_StructuredEventFormat asserts the JSON wire shape and ≥6 distinct phases
• Live smoke (build → :37788 → POST upload → SSE stream): 11 distinct phases, done:true received, server cleanly stopped

Caveats

• Vitest covers parseSseChunk (frame parsing) but not the full streaming hook against a mock ReadableStream; deferred since the parser is the load-bearing part and the live smoke covers wiring.
• Job's per-file event log retains the full history in memory (cleared on terminal). For very large multi-file uploads this could grow; not a regression vs. today's behaviour and acceptable for the typical UI use case.

🤖 Generated with Claude Code