feat: local LLM template metadata + refresh template store (v2.50.0) (czlonkowski#766)

Author: czlonkowski

CHANGELOG.md

@@ -7,6 +7,27 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [2.50.0] - 2026-05-02
+
+### Added
+
+- **Local LLM support for template metadata generation.** `fetch:templates --metadata-only` now routes to any OpenAI-compatible server (vLLM, Ollama, llama.cpp's `/v1`) when `N8N_MCP_LLM_BASE_URL` is set, falling back to OpenAI's Batch API otherwise. New `SequentialMetadataProcessor` issues direct `chat.completions.create` calls with configurable concurrency, since vLLM and friends do not implement OpenAI's `/v1/batches` endpoint. New env vars: `N8N_MCP_LLM_BASE_URL`, `N8N_MCP_LLM_MODEL` (default `Qwen/Qwen3.5-9B`), `N8N_MCP_LLM_API_KEY` (defaults to `EMPTY` for keyless local servers), `N8N_MCP_LLM_CONCURRENCY` (default 40). The cloud Batch path is unchanged.
+- **Stronger, leak-resistant prompt** for template metadata. The system message now spells out what each schema field means (categories, use_cases, required_services, key_features, target_audience) and explicitly forbids echoing prompt headers, which fixes a class of failures where smaller open-source models occasionally emitted `Template: ...` strings into the `categories` array. `createBatchRequest()` now delegates to `buildChatRequest()`, so the cloud Batch path picks up the new prompt too — both paths share the same body verbatim.
+
+### Changed
+
+- **Template store refreshed from n8n.io.** The templates table was rebuilt against the current API: 2,352 templates, 156 ranked node configurations across the most-popular nodes. Previous rebuild dated 2025-12-24.
+- **Template metadata regenerated end-to-end** against a local Qwen3.5-9B vLLM instance: 2,351/2,352 templates carry fresh `metadata_json` (99.96% coverage). One template (4334) skipped due to a tokenizer encoding edge case in its source content.
+- **Community node store refreshed** from the n8n Strapi verified list and the top-100 npm packages: **830 community nodes** (was 768, +62 new). Existing READMEs and AI summaries preserved through the upsert. Total nodes in DB: 1,650 (820 base + 830 community).
+- **Community AI documentation summaries regenerated** against the same local Qwen3.5-9B instance: **825/830 nodes** with both `npm_readme` and `ai_documentation_summary` (99.4% coverage). The 5 misses are npm packages that publish no README on npmjs, so there is no source text to summarise.
+
+### Notes
+
+- Template fetch only drops the `templates` and `templates_fts` tables — never `nodes`. Community nodes were verified intact at 768 mid-run before the separate community refresh added the 62 new ones.
+- A backup of the pre-fetch database lives at `/tmp/nodes-pre-template-update-20260502-093230.db` on the maintainer's machine.
+
+Conceived by Romuald Członkowski - https://www.aiadvisors.pl/en
+
 ## [2.49.0] - 2026-04-28
 
 ### Changed

README.md

@@ -9,19 +9,19 @@
 [![Docker](https://img.shields.io/badge/docker-ghcr.io%2Fczlonkowski%2Fn8n--mcp-green.svg)](https://github.com/czlonkowski/n8n-mcp/pkgs/container/n8n-mcp)
 [![Deploy on Railway](https://railway.com/button.svg)](https://railway.com/deploy/n8n-mcp?referralCode=n8n-mcp)
 
-A Model Context Protocol (MCP) server that provides AI assistants with comprehensive access to n8n node documentation, properties, and operations. Deploy in minutes to give Claude and other AI assistants deep knowledge about n8n's 1,588 workflow automation nodes (820 core + 768 community).
+A Model Context Protocol (MCP) server that provides AI assistants with comprehensive access to n8n node documentation, properties, and operations. Deploy in minutes to give Claude and other AI assistants deep knowledge about n8n's 1,650 workflow automation nodes (820 core + 830 community).
 
 ## Overview
 
 n8n-MCP serves as a bridge between n8n's workflow automation platform and AI models, enabling them to understand and work with n8n nodes effectively. It provides structured access to:
 
-- **1,505 n8n nodes** - 812 core nodes + 693 community nodes (605 verified)
+- **1,650 n8n nodes** - 820 core nodes + 830 community nodes (741 verified)
 - **Node properties** - 99% coverage with detailed schemas
 - **Node operations** - 63.6% coverage of available actions
 - **Documentation** - 87% coverage from official n8n docs (including AI nodes)
 - **AI tools** - 265 AI-capable tool variants detected with full documentation
-- **Real-world examples** - 2,646 pre-extracted configurations from popular templates
-- **Template library** - 2,709 workflow templates with 100% metadata coverage
+- **Real-world examples** - 156 ranked configurations extracted from popular templates
+- **Template library** - 2,352 workflow templates with 99.96% AI metadata coverage
 - **Community nodes** - Search verified community integrations with `source` filter
 
 ## Support This Project
@@ -102,7 +102,7 @@ CRITICAL: Execute tools without commentary. Only respond AFTER all tools complet
 When operations are independent, execute them in parallel for maximum performance.
 
 ### 3. Templates First
-ALWAYS check templates before building from scratch (2,709 available).
+ALWAYS check templates before building from scratch (2,352 available).
 
 ### 4. Multi-Level Validation
 Use validate_node(mode='minimal') → validate_node(mode='full') → validate_workflow pattern.
@@ -298,7 +298,7 @@ Use the same four-parameter format:
 ### Core Behavior
 1. **Silent execution** - No commentary between tools
 2. **Parallel by default** - Execute independent operations simultaneously
-3. **Templates first** - Always check before building (2,709 available)
+3. **Templates first** - Always check before building (2,352 available)
 4. **Multi-level validation** - Quick check → Full validation → Workflow validation
 5. **Never trust defaults** - Explicitly configure ALL parameters