Gemini response: Forge ecosystem consultation — hybrid search, plugin verdicts, wild ideas

Key findings:
- RAG fix: hybrid search (BM25 + vector), better embedding model, summary index
- Gitea plugin: recursive tree walker for .md files, confirmed approach
- RAGFlow: HOLD. R2: DEPLOY. n8n: CRUCIAL.
- Wild ideas: Awakened Concierge, Auto-Janitor, Jack Alert Override

Claude (Chronicler #82)

docs/consultations/gemini-forge-ecosystem-2026-04-12-response.md

@@ -0,0 +1,51 @@
+# Gemini Response: The Forge Ecosystem — April 12, 2026
+
+**Received:** April 12, 2026, ~3:30 AM CDT
+**From:** Gemini
+**Re:** The Forge RAG quality, plugin strategy, wild ideas
+
+---
+
+## Key Findings
+
+### RAG Problem — Root Cause Identified
+- **Not a missing data problem — it's a vector space problem**
+- `nomic-embed-text` maps "Emissary" near "diplomat/messenger/ambassador", not near "Meg"
+- Pure vector search can't bridge proper nouns to their context without explicit keyword matching
+
+### Fix Strategy (3 Steps)
+1. **Enable Hybrid Search** (BM25 + vector) — "the silver bullet"
+   - Switch from Vector Search to Hybrid Search in Dify Knowledge Base settings
+   - BM25 catches exact keywords, vectors catch semantic context
+2. **Upgrade Embedding Model** — `bge-m3` or `snowflake-arctic-embed-m`
+   - Both optimized for RAG, run well on CPU via Ollama
+   - Re-embed all 114 docs after switching
+3. **Use Summary Index** (Dify 1.12.0 feature)
+   - Generates document-level summaries for routing before chunk search
+   - FIREFROST-QUICK-REFERENCE.md would be instantly matched
+
+### Gitea Plugin Strategy — Confirmed Right Approach
+- Fork GitLab plugin, NOT just READMEs — need recursive tree walker
+- Key endpoints:
+  - Tree: `GET /api/v1/repos/:owner/:repo/git/trees/:branch?recursive=true`
+  - Filter: Only `.md` files, optionally only `docs/` directory
+  - Content: `GET /api/v1/repos/:owner/:repo/raw/:filepath`
+- Filter OUT code files — syntax noise destroys vector space
+
+### Plugin Ecosystem Verdicts
+| Plugin | Verdict | Reasoning |
+|--------|---------|-----------|
+| RAGFlow | **HOLD** | Dify + Hybrid Search sufficient for markdown. RAGFlow adds CPU load. |
+| Cloudflare R2 | **DEPLOY** | Artifact storage, The Forge's memory bank. Free egress. |
+| n8n Webhooks | **CRUCIAL** | Bridge between Dify output and infrastructure execution. |
+
+### Wild Card Ideas 🃏
+1. **Awakened Concierge** — n8n catches $1 payment webhook → Gemma 4 writes personalized welcome message in character → auto-posts to Discord
+2. **Pterodactyl Auto-Janitor** — Server crash → n8n sends last 50 console lines to Gemma 4 → AI diagnoses mod issue → suggests fix in Trinity Console for one-click approval
+3. **Jack Alert System Override** — Physical button/phone webhook → Discord AFK message → n8n suspends non-critical crons → infrastructure stabilizes automatically
+
+---
+
+## Gemini's Full Response
+
+[Full response preserved in session context — copy from Michael's Gemini chat if needed for verbatim archive]