Problem
The web UI already has AI features (summarize, auto-tag, suggest-links, RAG chat sidebar), but the launch plan calls for the web UI to be a full AI interaction portal where users can "see the knowledge graph grow as agents populate a KB" and "interact with the KB and workflows from BYOK model integration." Need to identify what's missing between current AI integration and the launch vision.
Audit Results (March 2026)
What Works — Full BYOK Support
All core AI features work end-to-end with user-provided API keys. No hardcoded keys, no forced cloud integrations.
| Feature | Status | Providers | Notes | |---------|--------|-----------|-------| | AI Provider Settings | Done | All 4 | Settings UI stores to DB, env vars as fallback | | Summarize | Done | All 4 | POST /api/ai/summarize, 256 token output | | Auto-Tag | Done | All 4 | Suggests 3-7 tags with reasoning, knows existing vocabulary | | Suggest Links | Done | All 4 | Hybrid search → AI evaluation of candidates | | RAG Chat Sidebar | Done | All 4 | SSE streaming, KB context, source citations, Cmd+Shift+K | | QA Dashboard | Done | N/A | Validation status, issue trends, entries needing review | | Semantic Search | Done | Local | sentence-transformers (all-MiniLM-L6-v2), no API key needed | | MCP Server | Done | N/A | Agents bring their own LLM, independent of Pyrite AI config |
Supported providers: Anthropic, OpenAI, OpenRouter, Ollama. All configured via Settings UI or env vars. DB settings take precedence over config.yaml.
Minor Issues (not blockers)
1. Test Connection is shallow — `/api/ai/status` only checks if key exists, doesn't validate it works. User gets "configured" with an invalid key, then errors on first AI operation. Fix: make test button do an actual API call. Effort: XS
2. Anthropic can't embed — Anthropic has no public embeddings API. If user picks Anthropic as their LLM provider, chat/summarize/tags work fine, but the embedding service uses local sentence-transformers anyway (independent of LLM provider). This is actually fine — embeddings are always local. No user-facing issue. No action needed.
3. Semantic search requires optional dep — `pip install 'pyrite[semantic]'` needed for sqlite-vec + sentence-transformers. Without it, search falls back to keyword silently. Docker image includes it by default. No action needed for deployed instances.
Remaining Gaps — Prioritized
Gaps from the original backlog item, re-assessed against current state:
| # | Gap | Effort | Priority | Milestone | Assessment | |---|-----|--------|----------|-----------|------------| | 1 | Gemini provider | M | Medium | 0.16+ | Separate backlog item gemini-byok-integration. Google AI SDK + Function Calling mapping. Not a blocker — 4 providers already cover most users. | | 2 | AI-assisted entry creation | S | Medium | 0.16+ | "Describe what you want" → structured entry. Nice UX but not blocking launch. Chat sidebar already helps with this workflow manually. | | 3 | Batch AI operations | S | Low | Post-launch | Summarize/tag/link-suggest across multiple entries. Useful for bulk imports but not a launch requirement. | | 4 | Live graph growth | M | Low | Post-launch | WebSocket infra exists but graph doesn't subscribe to live updates. Cool demo but not functional blocker. | | 5 | Workflow triggering from UI | L | Low | Post-launch | Start research/investigation workflows from browser. Requires agent orchestration layer that doesn't exist yet. | | 6 | Task monitoring dashboard | M | Low | Post-launch | Depends on task plugin (phases 1-2 done). Better suited for 0.17+ when task plugin matures. | | 7 | Multi-model chat | S | Low | Post-launch | Switch providers mid-conversation. Niche feature, post-launch. |
Verdict
BYOK is launch-ready. All four providers work. All AI features (summarize, tag, suggest-links, RAG chat) work with user-provided keys. Semantic search works locally with no API key. The remaining gaps are nice-to-have features for post-launch, not BYOK integration issues.
The only pre-launch improvement worth doing is the shallow test-connection fix (XS effort) — everything else can wait for 0.16+.
Deployment Notes for BYOK
Self-hosters: Configure AI in the Settings UI after first login. No env vars needed — the UI stores keys to the database.
Docker deployments: Optionally set env vars for server-wide defaults: ```bash ANTHROPIC_API_KEY=sk-ant-... # or OPENAI_API_KEY ``` Users can still override in Settings UI (DB settings take precedence).
Semantic search: Included in Docker image by default (`pyrite[semantic]`). Works without any API key — uses local sentence-transformers model.