spikerj/spikersoft-issues
1
0
Fork 0
Code Issues 196 Pull Requests Actions Packages Projects Releases Wiki Activity

[Mac local stack] Full site flow on the M5 — quiz generation + embeddings + book upload pipeline in mac-dev-up.sh #464

New Issue
Closed
opened 2026-07-08 23:43:09 +00:00 by spikerj · 4 comments
spikerj commented 2026-07-08 23:43:09 +00:00
Owner
Copy Link
Copy Source

Follow-up to #359 (closed; epic #346). #359 delivered the Art Studio flow on the Mac: scripts/mac-dev-up.sh brings up compose infra + native API + native ArtPipeProcessor (MPS). This ticket extends that to the entire site flow, so quiz generation can be tested locally alongside Art Studio. All work is in spikersoft-backend unless noted.

Why each piece is needed (traced, not speculative)

  • SpikerSoft.EventHandlers.QuizGeneration reads pre-extracted page text from Mongo (QuizGenerationWorkerService — "no PDF re-parsing needed"). Without the upload pipeline having run, there is nothing to generate quizzes from. The producer chain is: upload-coordinator → security-scanner (ClamAV) → metadata-extractor → book-management / file-movement, with SpikerSoft.EventHandlers.Embeddings storing pages + vectors (BookPageStorageService).
  • AI:EnableRAG defaults to true and queries RediSearch KNN (RedisVectorSearchService), which needs the embeddings worker to have populated the index. The mac compose override's single-node Redis Stack already loads the redisearch module.
  • Both quiz + embeddings csprojs hard-reference LLamaSharp.Backend.Cuda12, which cannot load on macOS.

Task 1 — LLamaSharp Metal backend (only real code change)

In SpikerSoft.EventHandlers.QuizGeneration/SpikerSoft.EventHandlers.QuizGeneration.csproj and SpikerSoft.EventHandlers.Embeddings/SpikerSoft.EventHandlers.Embeddings.csproj, make the backend package conditional:

<PackageReference Include="LLamaSharp.Backend.Cuda12" Version="0.27.0" Condition="!$([MSBuild]::IsOsPlatform('OSX'))" />
<PackageReference Include="LLamaSharp.Backend.Cpu" Version="0.27.0" Condition="$([MSBuild]::IsOsPlatform('OSX'))" />

Backend.Cpu ships Metal-enabled llama.cpp binaries for osx-arm64. Keep the existing comment about the PrivateAssets defense-in-depth. Verify Metal engages at runtime (llama.cpp logs ggml_metal / offloaded layers with GpuLayerCount=99). Note: this changes the Linux/CUDA build not at all — Docker images build on Linux and keep Cuda12.

Task 2 — GGUF models on the host

Add scripts/mac-dev-models.sh (idempotent, curl/hf download into ~/ai-models/gguf):

  • gemma-3n-E2B-it-Q8_0.gguf (Fast preset, ~5 GB)
  • gemma-3-12b-it-Q4_K_M.gguf (Balanced preset — default, ~9 GB)
  • nomic-embed-text-v2-moe.f16.gguf (embeddings)
  • Phi-4-reasoning-plus-Q4_K_M.gguf (Quality preset) optional / --all

Appsettings paths are container paths (/app/ai/...), so the native workers must be launched with AI__Models__Fast__Path, AI__Models__Balanced__Path, AI__Models__Quality__Path, and AI__EmbeddingModelPath overrides pointing at the host files.

Task 3 — Upload pipeline containers in the mac compose override

Add to docker-compose.mac.yml and to COMPOSE_SERVICES in scripts/mac-dev-common.sh: upload-coordinator, security-scanner, metadata-extractor, book-management, file-movement, clamav.

  • clamav: add platform: linux/amd64 (no arm64 image; runs under Rosetta — this was the documented workaround in the mac.yml header, now wired in for real). First boot downloads signatures (start_period 300s already in base file).
  • The base file's volume mounts are Windows paths (C:\temp\spikersoft-api-local\...). Override them in the mac file to a shared host dir, e.g. ~/spikersoft-local/{uploads,ebooks,blogs,dlls,code,assets,quarantine}.
  • The host-side API launched by mac-dev-up.sh must write uploads into that same shared dir — add the corresponding path config overrides to the api_env block (find the upload/asset root keys in SpikerSoft.Api appsettings).

Task 4 — Native quiz + embeddings workers in mac-dev-up.sh

New flag --full (implies existing behavior plus):

  • Build + launch SpikerSoft.EventHandlers.QuizGeneration and SpikerSoft.EventHandlers.Embeddings host-side, same nohup/pid-file/log pattern as the API and worker (.mac-dev/quiz.pid, .mac-dev/embeddings.pid, logs in .mac-dev/logs/).
  • Env for both: ASPNETCORE_ENVIRONMENT=Development, Mongo__ConnectionString=mongodb://localhost:27117/..., RabbitMQ__HostName=localhost, Redis__ConnectionString=localhost:6380,... (embeddings), Seq__ServerUrl=http://localhost:5341, Jaeger__EndPoint=http://localhost:4317, GpuScheduling__Mode=dedicated (single-user laptop; no coordinator arbitration needed), model path overrides from Task 2, distinct Kestrel ports for healthz (avoid 8080 — the art worker owns it; suggest 8180/8181).
  • Preflight: warn when GGUFs are missing (mirror the venv warning pattern) and point at scripts/mac-dev-models.sh.
  • scripts/mac-dev-down.sh stops both.

Task 5 — End-to-end smoke

Add scripts/mac-dev-smoke-quiz.sh (or extend mac-dev-smoke.sh): upload a small PDF via the API (HTTPS :5291, multipart), poll Mongo until pre-extracted pages exist for the book, trigger quiz generation via the API, poll until the quiz document appears; print timings. If the RAG/RediSearch path misbehaves on single-node Redis Stack, fall back to AI__EnableRAG=false for the quiz worker and note it in the doc.

Task 6 — Docs

Update docs/mac-local-stack.md: full-flow recipe (mac-dev-models.sh once → mac-dev-up.sh --full), download sizes, shared-dir layout, ClamAV/Rosetta caveat, healthz port map.

Out of scope

  • image-description (Qwen3-VL) on MPS — possible natively via transformers, not needed for quiz or art flows; separate ticket if wanted.
  • CUDA-only art lanes (TripoSG, Hunyuan3DPaint) — no Mac path exists; production hardware track covers these.

Acceptance

On the M5 with venvs + Blender already set up per #359: scripts/mac-dev-models.sh && scripts/mac-dev-up.sh --full brings everything up; uploading a PDF through the local site produces an available book; requesting a quiz on it returns generated questions using Metal-accelerated gemma (verify offload in the quiz log); the Art Studio smoke from #359 still passes; mac-dev-down.sh --all tears down clean.

Follow-up to #359 (closed; epic #346). #359 delivered the Art Studio flow on the Mac: `scripts/mac-dev-up.sh` brings up compose infra + native API + native ArtPipeProcessor (MPS). This ticket extends that to the **entire site flow**, so quiz generation can be tested locally alongside Art Studio. All work is in `spikersoft-backend` unless noted. ## Why each piece is needed (traced, not speculative) - `SpikerSoft.EventHandlers.QuizGeneration` reads **pre-extracted page text from Mongo** (`QuizGenerationWorkerService` — "no PDF re-parsing needed"). Without the upload pipeline having run, there is nothing to generate quizzes from. The producer chain is: upload-coordinator → security-scanner (ClamAV) → metadata-extractor → book-management / file-movement, with `SpikerSoft.EventHandlers.Embeddings` storing pages + vectors (`BookPageStorageService`). - `AI:EnableRAG` defaults to `true` and queries RediSearch KNN (`RedisVectorSearchService`), which needs the embeddings worker to have populated the index. The mac compose override's single-node Redis Stack already loads the redisearch module. - Both quiz + embeddings csprojs hard-reference `LLamaSharp.Backend.Cuda12`, which cannot load on macOS. ## Task 1 — LLamaSharp Metal backend (only real code change) In `SpikerSoft.EventHandlers.QuizGeneration/SpikerSoft.EventHandlers.QuizGeneration.csproj` and `SpikerSoft.EventHandlers.Embeddings/SpikerSoft.EventHandlers.Embeddings.csproj`, make the backend package conditional: ```xml <PackageReference Include="LLamaSharp.Backend.Cuda12" Version="0.27.0" Condition="!$([MSBuild]::IsOsPlatform('OSX'))" /> <PackageReference Include="LLamaSharp.Backend.Cpu" Version="0.27.0" Condition="$([MSBuild]::IsOsPlatform('OSX'))" /> ``` `Backend.Cpu` ships Metal-enabled llama.cpp binaries for osx-arm64. Keep the existing comment about the PrivateAssets defense-in-depth. Verify Metal engages at runtime (llama.cpp logs `ggml_metal` / offloaded layers with `GpuLayerCount=99`). Note: this changes the Linux/CUDA build not at all — Docker images build on Linux and keep Cuda12. ## Task 2 — GGUF models on the host Add `scripts/mac-dev-models.sh` (idempotent, curl/hf download into `~/ai-models/gguf`): - `gemma-3n-E2B-it-Q8_0.gguf` (Fast preset, ~5 GB) - `gemma-3-12b-it-Q4_K_M.gguf` (Balanced preset — default, ~9 GB) - `nomic-embed-text-v2-moe.f16.gguf` (embeddings) - `Phi-4-reasoning-plus-Q4_K_M.gguf` (Quality preset) optional / `--all` Appsettings paths are container paths (`/app/ai/...`), so the native workers must be launched with `AI__Models__Fast__Path`, `AI__Models__Balanced__Path`, `AI__Models__Quality__Path`, and `AI__EmbeddingModelPath` overrides pointing at the host files. ## Task 3 — Upload pipeline containers in the mac compose override Add to `docker-compose.mac.yml` and to `COMPOSE_SERVICES` in `scripts/mac-dev-common.sh`: `upload-coordinator`, `security-scanner`, `metadata-extractor`, `book-management`, `file-movement`, `clamav`. - `clamav`: add `platform: linux/amd64` (no arm64 image; runs under Rosetta — this was the documented workaround in the mac.yml header, now wired in for real). First boot downloads signatures (start_period 300s already in base file). - The base file's volume mounts are Windows paths (`C:\temp\spikersoft-api-local\...`). Override them in the mac file to a shared host dir, e.g. `~/spikersoft-local/{uploads,ebooks,blogs,dlls,code,assets,quarantine}`. - The host-side API launched by mac-dev-up.sh must write uploads into that same shared dir — add the corresponding path config overrides to the `api_env` block (find the upload/asset root keys in `SpikerSoft.Api` appsettings). ## Task 4 — Native quiz + embeddings workers in mac-dev-up.sh New flag `--full` (implies existing behavior plus): - Build + launch `SpikerSoft.EventHandlers.QuizGeneration` and `SpikerSoft.EventHandlers.Embeddings` host-side, same nohup/pid-file/log pattern as the API and worker (`.mac-dev/quiz.pid`, `.mac-dev/embeddings.pid`, logs in `.mac-dev/logs/`). - Env for both: `ASPNETCORE_ENVIRONMENT=Development`, `Mongo__ConnectionString=mongodb://localhost:27117/...`, `RabbitMQ__HostName=localhost`, `Redis__ConnectionString=localhost:6380,...` (embeddings), `Seq__ServerUrl=http://localhost:5341`, `Jaeger__EndPoint=http://localhost:4317`, `GpuScheduling__Mode=dedicated` (single-user laptop; no coordinator arbitration needed), model path overrides from Task 2, distinct Kestrel ports for healthz (avoid 8080 — the art worker owns it; suggest 8180/8181). - Preflight: warn when GGUFs are missing (mirror the venv warning pattern) and point at `scripts/mac-dev-models.sh`. - `scripts/mac-dev-down.sh` stops both. ## Task 5 — End-to-end smoke Add `scripts/mac-dev-smoke-quiz.sh` (or extend mac-dev-smoke.sh): upload a small PDF via the API (HTTPS :5291, multipart), poll Mongo until pre-extracted pages exist for the book, trigger quiz generation via the API, poll until the quiz document appears; print timings. If the RAG/RediSearch path misbehaves on single-node Redis Stack, fall back to `AI__EnableRAG=false` for the quiz worker and note it in the doc. ## Task 6 — Docs Update `docs/mac-local-stack.md`: full-flow recipe (`mac-dev-models.sh` once → `mac-dev-up.sh --full`), download sizes, shared-dir layout, ClamAV/Rosetta caveat, healthz port map. ## Out of scope - image-description (Qwen3-VL) on MPS — possible natively via transformers, not needed for quiz or art flows; separate ticket if wanted. - CUDA-only art lanes (TripoSG, Hunyuan3DPaint) — no Mac path exists; production hardware track covers these. ## Acceptance On the M5 with venvs + Blender already set up per #359: `scripts/mac-dev-models.sh && scripts/mac-dev-up.sh --full` brings everything up; uploading a PDF through the local site produces an available book; requesting a quiz on it returns generated questions using Metal-accelerated gemma (verify offload in the quiz log); the Art Studio smoke from #359 still passes; `mac-dev-down.sh --all` tears down clean.
spikerj added the enhancement label 2026-07-08 23:43:27 +00:00
spikerj commented 2026-07-09 00:38:12 +00:00
Author
Owner
Copy Link
Copy Source

Status 2026-07-08: implementation complete, committed locally on spikersoft-backend branch feat/464-mac-full-site-flow (add055b). Paused before the full acceptance run — home internet is currently unusable (~200 KB/s, flaky), which blocks the GGUF downloads and the in-Docker NuGet restores.

Done & verified on the M5

  • Task 1 — LLamaSharp.Backend.Cuda12 is now Condition="!IsOsPlatform('OSX')" with Backend.Cpu on OSX in both csprojs. Both workers restore/build on macOS and runtimes/osx-arm64/native/libggml-metal.dylib is in the output (Metal binaries confirmed). Linux/Docker builds unchanged.
  • Task 2 — scripts/mac-dev-models.sh (idempotent, curl resume, --all for Phi-4). Note: the README's google//bartowski/ HF URLs are gated or 404 today; switched to the ungated unsloth/ + nomic-ai/ mirrors (exact filename matches verified via HF API; README updated).
  • Task 3 — pipeline services in docker-compose.mac.yml (clamav platform: linux/amd64), Windows mounts replaced with $MAC_SHARED_DIR (~/spikersoft-local). Design note: the pipeline messages carry absolute StagingPaths and metadata-extractor hardcodes /app/ebooks in code, so each container mounts the shared dir at both /app/<subdir> and the identical absolute host path (mirror mount). Merged compose config validates. Kept the pipeline services out of the base COMPOSE_SERVICES (art-only bring-up shouldn't build 5 images / run Rosetta ClamAV) — they're a separate UPLOAD_PIPELINE_SERVICES list appended by --full.
  • Task 4 — mac-dev-up.sh --full builds+launches both native workers (pids/logs in .mac-dev/), GGUF preflight warning, health waits on :8180/:8181; mac-dev-down.sh stops them (exercised). Extra code change the ticket needed: WithKestrel now honors a Kestrel:Port config override — it calls UseUrls, which env vars can't override, so 8180/8181 was impossible without it. Containers don't set the key; deployed behavior unchanged. Also added MAC_DEV_ENABLE_RAG=false as the RediSearch fallback knob.
  • Task 5 — scripts/mac-dev-smoke-quiz.sh: generates a text-only PDF via cupsfilter (books with images wait ~30 min on the out-of-scope image-description worker — documented), uploads as test user, approves as the seeded e2e.staff account, follows the workflow in Mongo (status → book-pages → EmbeddingsComplete → quiz doc), prints timings + sample question. Second extra code change: ApproveUpload requires the staff role, which the role-less test user lacks — under UseTestAuthentication=true the real Keycloak JWT bearer scheme now stays registered alongside the test scheme. Verified live on the M5: token-less → 403 on GET /api/book/pending (art-studio behavior from #359 unchanged), real e2e.staff token (password grant, spikersoft-backend client) → 200. Keycloak only contacted when a Bearer token is present, so the stack stays offline-capable. Bonus: a locally-served Angular app can now hit the local API with real logins.
  • Task 6 — docs/mac-local-stack.md §7: recipe, model table with sizes/sources, healthz port map (API 5290/5291, art 8080, quiz 8180, embeddings 8181), shared-dir layout, ClamAV/Rosetta + GET /api/book/{id}/file caveats.
  • SpikerSoft.Tests.Unit: 11437 passed, 4 failed — the identical 4 fail on unmodified master (environmental: GitService per-user repo ×2, Python curriculum ×2). shellcheck clean.

Blocked on connectivity (resume when internet is good)

  1. scripts/mac-dev-models.sh — resumes the three .part files already in ~/ai-models/gguf (downloads smallest-first: nomic → gemma-3n → gemma-12b).
  2. docker compose -f docker-compose.yml -f docker-compose.mac.yml -f docker-compose.redisport.yml build upload-coordinator security-scanner metadata-extractor book-management file-movement (first attempt died in in-container dotnet restore) — or just let mac-dev-up.sh --full build them.
  3. scripts/mac-dev-up.sh --full → scripts/mac-dev-smoke-quiz.sh (needs spikersoft-angular/.env.e2e staff creds — already present on the M5), verify ggml_metal offload in .mac-dev/logs/quiz.log, re-run the #359 art smoke, mac-dev-down.sh --all.
  4. Push branch + open the PR once the smoke passes.
**Status 2026-07-08: implementation complete, committed locally on `spikersoft-backend` branch `feat/464-mac-full-site-flow` (add055b). Paused before the full acceptance run — home internet is currently unusable (~200 KB/s, flaky), which blocks the GGUF downloads and the in-Docker NuGet restores.** ### Done & verified on the M5 - **Task 1** — `LLamaSharp.Backend.Cuda12` is now `Condition="!IsOsPlatform('OSX')"` with `Backend.Cpu` on OSX in both csprojs. Both workers restore/build on macOS and `runtimes/osx-arm64/native/libggml-metal.dylib` is in the output (Metal binaries confirmed). Linux/Docker builds unchanged. - **Task 2** — `scripts/mac-dev-models.sh` (idempotent, curl resume, `--all` for Phi-4). Note: the README's `google/`/`bartowski/` HF URLs are gated or 404 today; switched to the ungated `unsloth/` + `nomic-ai/` mirrors (exact filename matches verified via HF API; README updated). - **Task 3** — pipeline services in `docker-compose.mac.yml` (clamav `platform: linux/amd64`), Windows mounts replaced with `$MAC_SHARED_DIR` (~/spikersoft-local). Design note: the pipeline messages carry **absolute** StagingPaths and metadata-extractor hardcodes `/app/ebooks` in code, so each container mounts the shared dir at both `/app/<subdir>` and the identical absolute host path (mirror mount). Merged compose config validates. Kept the pipeline services out of the base `COMPOSE_SERVICES` (art-only bring-up shouldn't build 5 images / run Rosetta ClamAV) — they're a separate `UPLOAD_PIPELINE_SERVICES` list appended by `--full`. - **Task 4** — `mac-dev-up.sh --full` builds+launches both native workers (pids/logs in `.mac-dev/`), GGUF preflight warning, health waits on :8180/:8181; `mac-dev-down.sh` stops them (exercised). Extra code change the ticket needed: `WithKestrel` now honors a `Kestrel:Port` config override — it calls `UseUrls`, which env vars can't override, so 8180/8181 was impossible without it. Containers don't set the key; deployed behavior unchanged. Also added `MAC_DEV_ENABLE_RAG=false` as the RediSearch fallback knob. - **Task 5** — `scripts/mac-dev-smoke-quiz.sh`: generates a **text-only** PDF via cupsfilter (books with images wait ~30 min on the out-of-scope image-description worker — documented), uploads as test user, approves as the seeded `e2e.staff` account, follows the workflow in Mongo (status → book-pages → EmbeddingsComplete → quiz doc), prints timings + sample question. **Second extra code change:** `ApproveUpload` requires the staff role, which the role-less test user lacks — under `UseTestAuthentication=true` the real Keycloak JWT bearer scheme now stays registered alongside the test scheme. Verified live on the M5: token-less → 403 on `GET /api/book/pending` (art-studio behavior from #359 unchanged), real `e2e.staff` token (password grant, `spikersoft-backend` client) → 200. Keycloak only contacted when a Bearer token is present, so the stack stays offline-capable. Bonus: a locally-served Angular app can now hit the local API with real logins. - **Task 6** — `docs/mac-local-stack.md` §7: recipe, model table with sizes/sources, healthz port map (API 5290/5291, art 8080, quiz 8180, embeddings 8181), shared-dir layout, ClamAV/Rosetta + `GET /api/book/{id}/file` caveats. - `SpikerSoft.Tests.Unit`: 11437 passed, 4 failed — the identical 4 fail on unmodified master (environmental: GitService per-user repo ×2, Python curriculum ×2). shellcheck clean. ### Blocked on connectivity (resume when internet is good) 1. `scripts/mac-dev-models.sh` — resumes the three `.part` files already in `~/ai-models/gguf` (downloads smallest-first: nomic → gemma-3n → gemma-12b). 2. `docker compose -f docker-compose.yml -f docker-compose.mac.yml -f docker-compose.redisport.yml build upload-coordinator security-scanner metadata-extractor book-management file-movement` (first attempt died in in-container `dotnet restore`) — or just let `mac-dev-up.sh --full` build them. 3. `scripts/mac-dev-up.sh --full` → `scripts/mac-dev-smoke-quiz.sh` (needs `spikersoft-angular/.env.e2e` staff creds — already present on the M5), verify `ggml_metal` offload in `.mac-dev/logs/quiz.log`, re-run the #359 art smoke, `mac-dev-down.sh --all`. 4. Push branch + open the PR once the smoke passes.
spikerj commented 2026-07-09 00:44:18 +00:00
Author
Owner
Copy Link
Copy Source

Task 1 done — merged to master in spikersoft-backend PR #188 (commit d0e65d8).

LLamaSharp.Backend.Cuda12 is now OS-conditional (!IsOsPlatform('OSX')) with LLamaSharp.Backend.Cpu added on OSX, in both SpikerSoft.EventHandlers.QuizGeneration.csproj and SpikerSoft.EventHandlers.Embeddings.csproj. Backend.Cpu ships Metal-enabled llama.cpp binaries for osx-arm64. The Linux/CUDA production build is unchanged (negated condition + Docker builds on Linux). Verified locally on macOS arm64: QuizGeneration, Embeddings, and SpikerSoft.Tests.Unit all build 0 errors.

Leaving this ticket open — Tasks 2–6 (scripts/mac-dev-models.sh, the mac compose upload-pipeline services, the --full native quiz+embeddings launch in mac-dev-up.sh, the E2E smoke script, and docs) remain, and are in progress on branch feat/464-mac-full-site-flow. Runtime Metal-offload verification depends on the Task 2 GGUFs.

**Task 1 done** — merged to `master` in spikersoft-backend PR #188 (commit `d0e65d8`). `LLamaSharp.Backend.Cuda12` is now OS-conditional (`!IsOsPlatform('OSX')`) with `LLamaSharp.Backend.Cpu` added on OSX, in both `SpikerSoft.EventHandlers.QuizGeneration.csproj` and `SpikerSoft.EventHandlers.Embeddings.csproj`. Backend.Cpu ships Metal-enabled llama.cpp binaries for osx-arm64. The Linux/CUDA production build is unchanged (negated condition + Docker builds on Linux). Verified locally on macOS arm64: QuizGeneration, Embeddings, and `SpikerSoft.Tests.Unit` all build 0 errors. Leaving this ticket **open** — Tasks 2–6 (`scripts/mac-dev-models.sh`, the mac compose upload-pipeline services, the `--full` native quiz+embeddings launch in `mac-dev-up.sh`, the E2E smoke script, and docs) remain, and are in progress on branch `feat/464-mac-full-site-flow`. Runtime Metal-offload verification depends on the Task 2 GGUFs.
spikerj commented 2026-07-09 09:56:43 +00:00
Author
Owner
Copy Link
Copy Source

Acceptance run complete (M5, overnight 2026-07-09) — everything passes. Branch feat/464-mac-full-site-flow (00f672d) pushed to spikersoft-backend; PR not yet opened (waiting for Joey's go-ahead).

  • mac-dev-models.sh → 3 GGUFs, 12.2 GB (resumed the .part files once the connection recovered; ~1 MB/s overnight)
  • mac-dev-up.sh --full → infra + 6 pipeline containers + API + art/quiz/embeddings workers healthy in ~21 s (after one-time image builds)
  • mac-dev-smoke-quiz.sh → PASSED: upload → e2e.staff approval (real Keycloak token against the test-auth API) → ClamAV scan t+4 s → 8 pre-extracted pages in book-pages t+8 s → embeddings + BookPageChunk:* RediSearch vectors t+18 s → 9 RAG-grounded questions t+114 s (sample: "According to the provided text, what is a primary function of message brokers like RabbitMQ?")
  • Metal offload confirmed: every gemma layer assigned to device MTL0 in quiz.log; ggml_metal_device_init: GPU family MTLGPUFamilyApple10
  • #359 art smoke still passes: Prop plan, 5 stages, 47 s, all artifacts in GridFS
  • mac-dev-down.sh --all tears down clean (no orphan processes/containers); restart-after-teardown verified — stack left running
  • One extra fix found during the run: the base compose ClamAV healthcheck (/usr/local/bin/clamd --ping) doesn't exist in current clamav images, so the container reported unhealthy forever — the mac override now probes with clamdcheck.sh (fix included in the branch)

Remaining: open the PR from feat/464-mac-full-site-flow → master, merge after CI, then comment+close this ticket per the tracker workflow.

**Acceptance run complete (M5, overnight 2026-07-09) — everything passes. Branch `feat/464-mac-full-site-flow` (00f672d) pushed to `spikersoft-backend`; PR not yet opened (waiting for Joey's go-ahead).** - `mac-dev-models.sh` → 3 GGUFs, 12.2 GB (resumed the `.part` files once the connection recovered; ~1 MB/s overnight) - `mac-dev-up.sh --full` → infra + 6 pipeline containers + API + art/quiz/embeddings workers healthy in ~21 s (after one-time image builds) - `mac-dev-smoke-quiz.sh` → **PASSED**: upload → `e2e.staff` approval (real Keycloak token against the test-auth API) → ClamAV scan t+4 s → 8 pre-extracted pages in `book-pages` t+8 s → embeddings + `BookPageChunk:*` RediSearch vectors t+18 s → **9 RAG-grounded questions t+114 s** (sample: "According to the provided text, what is a primary function of message brokers like RabbitMQ?") - Metal offload confirmed: every gemma layer `assigned to device MTL0` in quiz.log; `ggml_metal_device_init: GPU family MTLGPUFamilyApple10` - #359 art smoke still passes: Prop plan, 5 stages, 47 s, all artifacts in GridFS - `mac-dev-down.sh --all` tears down clean (no orphan processes/containers); restart-after-teardown verified — stack left running - One extra fix found during the run: the base compose ClamAV healthcheck (`/usr/local/bin/clamd --ping`) doesn't exist in current clamav images, so the container reported unhealthy forever — the mac override now probes with `clamdcheck.sh` (fix included in the branch) Remaining: open the PR from `feat/464-mac-full-site-flow` → master, merge after CI, then comment+close this ticket per the tracker workflow.
spikerj commented 2026-07-11 23:26:29 +00:00
Author
Owner
Copy Link
Copy Source

Resolved in spikersoft-backend PR #198 (merged; Task 1 in PR #188): mac-dev-up.sh --full brings up the entire site flow on the M5 — quiz generation + embeddings + book upload pipeline — with OS-conditional LLamaSharp Metal backend. Acceptance-verified 2026-07-09. Closing.

Resolved in spikersoft-backend PR #198 (merged; Task 1 in PR #188): mac-dev-up.sh --full brings up the entire site flow on the M5 — quiz generation + embeddings + book upload pipeline — with OS-conditional LLamaSharp Metal backend. Acceptance-verified 2026-07-09. Closing.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
agentic
Actionable surfaced by the agentic architecture & security review
 bug
Something is not working
 duplicate
This issue or pull request already exists
 enhancement
New feature
 help wanted
Need some help
 invalid
Something is wrong
 question
More information is needed
 sonarqube
Code quality issue surfaced by SonarQube (bug / code smell)
 wontfix
This won't be fixed
No labels enhancement
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
spikerj (Joseph Spiker)
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: spikerj/spikersoft-issues#464