Apply Q1–Q3 decisions: GraphDB/SPARQL, reuse Qdrant, self-host embeddings via TEI

- Q1 → Ontotext GraphDB (SPARQL). docs/08 fully rewritten with RDF schema,
  example SPARQL queries (transitive deps via property paths, time-windowed
  topic counts, cross-pillar joins via nexa:vectorId).
- Q2 → reuse qdrant_scientific with nexa_* collection prefix; docs/09 step 2
  now points there explicitly.
- Q3 → no OpenAI embeddings. Self-host on the docker host. Use TEI
  (HuggingFace text-embeddings-inference) — single Rust binary, ~500 MB image,
  OpenAI-compatible — instead of Ollama, since we only need embeddings.
- docs/09 Phase-3.4 add-on simplified to a single Ontotext compose snippet
  (Neo4j option dropped) plus repo creation curl.
- docs/11 Q3 marked resolved; new Q15 picks the model (bge-m3 vs nomic-embed)
  and adds the open question of whether SAIA already proxies an embedding
  model that would let us skip TEI entirely.
- docs/03 + CLAUDE.md updated with the new decisions so future runs don't
  re-litigate.

docs/11-open-questions.md

@@ -2,13 +2,22 @@
 
 Items that **block progress** and need a human decision before a workflow can be implemented or a service deployed. Tick them off as you decide.
 
+## Resolved
+
+- [x] **Q1 — Graph DB choice → Ontotext GraphDB (SPARQL).** Rationale: explore Nexa's memory through SPARQL is a stated goal. [08-graphrag-architecture](./08-graphrag-architecture.md) is rewritten accordingly.
+- [x] **Q2 — Vector store → reuse `qdrant_scientific`** with a `nexa_*` collection prefix. No dedicated container.
+- [x] **Q3 — Embeddings model → not OpenAI.** Self-host on the docker host via **TEI** (HF `text-embeddings-inference`) — Rust single-binary, OpenAI-compatible, ~500 MB image, no LLM runtime overhead. Speed analysis in §"Speed budget" below; concrete model still open as **Q15**.
+
 ## Architectural decisions
 
-- [ ] **Q1 — Graph DB choice.** Phase 3.4 says *Neo4j* in [08](./08-graphrag-architecture.md) but the original deployment draft used *Ontotext GraphDB*. **Recommendation:** Neo4j (smaller footprint, Cypher is friendlier for n8n HTTP nodes, no SHACL/OWL needs identified). Confirm or override.
-- [ ] **Q2 — Reuse `qdrant_scientific` or run a dedicated `nexa-qdrant`?** Reuse keeps memory usage flat (~670 MB seen on Homepage); dedicated gives clean isolation and independent snapshot lifecycle. **Recommendation:** reuse with a `nexa_*` collection prefix.
-- [ ] **Q3 — Embeddings model.** Default in `infra_manifest.yaml` is `text-embedding-3-small` (OpenAI). Is the LiteLLM virtual key authorised to call it, or should we route to a self-hosted alternative (e.g. `bge-m3` via Ollama on the same host)?
 - [ ] **Q4 — Obsidian sync mechanism.** `system_prime.txt` references Obsidian Context, but the current setup syncs via Nextcloud (`nc.nuclide.systems` → `Notizen` folder, ~200 MB). Should Nexa watch the **filesystem on LXC 105** (NC data dir) or **the Nextcloud WebDAV API**? FS is cheaper, WebDAV is portable.
-- [ ] **Q5 — Karakeep vs. Hoarder naming.** The Zoraxy host is `hoarder.nuclide.systems` but the container is `karakeep-*` and Homepage labels it Karakeep. They're the same product (rename in 2024). Pick one display name to use in docs and prompts.
+- [ ] **Q5 — Karakeep vs. Hoarder naming.** Zoraxy host is `hoarder.nuclide.systems` but containers are `karakeep-*` and Homepage labels it Karakeep. Same product (rename 2024). Pick one display name for docs and prompts.
+- [ ] **Q15 — Embedding model (served via TEI).** Two viable options on the docker host (CPU only — see speed budget):
+  - **`BAAI/bge-m3`** (568 M params, ~1 GB RAM, multilingual incl. German, 1024-dim, ~10–20 docs/s on CPU). **Recommended** because the corpus is DE/EN-mixed. Requires `qdrant_schema.json` `vectors.size = 1024`.
+  - **`nomic-ai/nomic-embed-text-v1.5`** (137 M, ~250 MB RAM, EN-leaning, 768-dim, ~50 docs/s CPU). Lighter / faster but weaker on German.
+  - **(maybe) does SAIA already proxy any embedding model?** If the SAIA backend offers e.g. `mistral-embed` for free, we can skip TEI entirely. Worth a 1-line check in the LiteLLM admin UI before deploying TEI.
+
+  Decide before Phase 3.1.
 
 ## Identifiers needed (auto-discoverable, but list now if known)
 
@@ -20,10 +29,25 @@ Items that **block progress** and need a human decision before a workflow can be
 
 ## Hardware / capacity
 
-- [ ] **Q11 — RAM headroom on docker host.** Homepage shows 29.5 GiB free out of ~31 GiB total, ~8.5 GB used. Phase-3 Qdrant indexing of Obsidian + Memos history is fine here, but a Phase-3.4 GraphDB choice changes this (Neo4j ~2 GB heap, Ontotext ~4 GB). Confirm acceptable.
-- [ ] **Q12 — S3 archive bucket.** `s3.nuclide.systems` is up. Bucket name + access key for Qdrant snapshots?
+- [ ] **Q11 — RAM headroom on docker host.** 29.5 GiB free / ~31 GiB total, ~8.5 GB used. Phase-3 Qdrant indexing + TEI/`bge-m3` (~1.1 GB) + Ontotext GraphDB (~4 GB heap) ⇒ ~14 GB used worst case, still ample. Confirm acceptable.
+- [ ] **Q12 — S3 archive bucket.** `s3.nuclide.systems` is up. Bucket name + access key for Qdrant snapshots and GraphDB exports?
 
 ## Process
 
-- [ ] **Q13 — Octoprint container is Exited** (Homepage). Out of scope for Nexa, but flagging it because Phase 5 monitoring would alert on it. Suppress, or is it intentional?
-- [ ] **Q14 — `Missing Widget Type: zoraxy`** on Homepage — Homepage doesn't have a Zoraxy widget yet. Cosmetic, unrelated to Nexa.
+- [ ] **Q13 — Octoprint container is Exited** (Homepage). Out of scope for Nexa, but Phase-5 monitoring would alert on it. Suppress or is it intentional?
+- [ ] **Q14 — `Missing Widget Type: zoraxy`** on Homepage. Cosmetic, unrelated to Nexa.
+
+---
+
+## Speed budget (Q3 follow-up)
+
+Workload on the docker host (16 CPU, ~30 GB free RAM):
+
+| Task | Volume | Latency target | Achievable on CPU with `bge-m3` | Achievable with `nomic-embed-text` |
+|------|--------|----------------|---------------------------------|------------------------------------|
+| Real-time memo embed | 1 doc | <500 ms incl. n8n round-trip | ✅ ~50–100 ms | ✅ ~20 ms |
+| Daily ingest | ~70 docs | <60 s | ✅ ~5–10 s | ✅ ~2 s |
+| Obsidian backfill (one-shot) | ~2 000 docs | <15 min | ✅ ~2–4 min | ✅ <1 min |
+| RAG query embed (`#nexa:ask`) | 1 doc | <300 ms | ✅ ~50 ms | ✅ ~20 ms |
+
+Conclusion: CPU-only Ollama is sufficient — no GPU needed for current scope. Bottleneck is SAIA chat (already remote), not embeddings.