How do I install MCP Local RAG?

Install MCP Local RAG with a single command: npx mdskills install shinpr/mcp-local-rag. This downloads the skill files into your project and your AI agent picks them up automatically.

What platforms support MCP Local RAG?

MCP Local RAG works with Claude Code, Claude Desktop, Cursor, Vscode Copilot, Windsurf, Continue Dev, Gemini Cli, Amp, Roo Code, Goose. Skills use the open SKILL.md format which is compatible with any AI coding agent that reads markdown instructions.

← Back to MCP servers

MCP Local RAG

Name: MCP Local RAG: AI Agent Skill
Rating: 8 (1 reviews)
Author: shinpr

Verified

MCP ServerIntermediate

Provides score interpretation (< 0.3 good, > 0.5 skip), query optimization, and source naming for query_documents, ingest_file, ingest_data tools. Use this skill when working with RAG, searching documents, ingesting files, saving web content, or handling PDF, HTML, DOCX, TXT, Markdown.

by @shinpr0Updated 1 weeks ago

Add this skill

npx mdskills install shinpr/mcp-local-rag

Fork & Edit

Skill Advisor8.0

Comprehensive RAG guidance with score thresholds, query strategies, and ingestion patterns

+Provides actionable score thresholds for filtering search results by relevance
+Defines clear query optimization rules for transforming vague or technical queries
+Covers both ingestion patterns with metadata guidance and result selection criteria
-References external documentation files that may not be accessible to agents

SKILL.md

Edit in Browser

1---
2name: mcp-local-rag
3description: Provides score interpretation (< 0.3 good, > 0.5 skip), query optimization, and source naming for query_documents, ingest_file, ingest_data tools. Use this skill when working with RAG, searching documents, ingesting files, saving web content, or handling PDF, HTML, DOCX, TXT, Markdown.
4---
5 
6# MCP Local RAG Skills
7 
8## Tools
9 
10| Tool | Use When |
11|------|----------|
12| `ingest_file` | Local files (PDF, DOCX, TXT, MD) |
13| `ingest_data` | Raw content (HTML, text) with source URL |
14| `query_documents` | Semantic + keyword hybrid search |
15| `delete_file` / `list_files` / `status` | Management |
16 
17## Search: Core Rules
18 
19Hybrid search combines vector (semantic) and keyword (BM25).
20 
21### Score Interpretation
22 
23Lower = better match. Use this to filter noise.
24 
25| Score | Action |
26|-------|--------|
27| < 0.3 | Use directly |
28| 0.3-0.5 | Include if mentions same concept/entity |
29| > 0.5 | Skip unless no better results |
30 
31### Limit Selection
32 
33| Intent | Limit |
34|--------|-------|
35| Specific answer (function, error) | 5 |
36| General understanding | 10 |
37| Comprehensive survey | 20 |
38 
39### Query Formulation
40 
41| Situation | Why Transform | Action |
42|-----------|---------------|--------|
43| Specific term mentioned | Keyword search needs exact match | KEEP term |
44| Vague query | Vector search needs semantic signal | ADD context |
45| Error stack or code block | Long text dilutes relevance | EXTRACT core keywords |
46| Multiple distinct topics | Single query conflates results | SPLIT queries |
47| Few/poor results | Term mismatch | EXPAND (see below) |
48 
49### Query Expansion
50 
51When results are few or all score > 0.5, expand query terms:
52 
53- Keep original term first, add 2-4 variants
54- Types: synonyms, abbreviations, related terms, word forms
55- Example: `"config"` → `"config configuration settings configure"`
56 
57Avoid over-expansion (causes topic drift).
58 
59### Result Selection
60 
61When to include vs skip—based on answer quality, not just score.
62 
63**INCLUDE** if:
64- Directly answers the question
65- Provides necessary context
66- Score < 0.5
67 
68**SKIP** if:
69- Same keyword, unrelated context
70- Score > 0.7
71- Mentions term without explanation
72 
73### fileTitle
74 
75Each result includes `fileTitle` (document title extracted from content). Null when extraction fails.
76 
77| Use | How |
78|-----|-----|
79| Disambiguate chunks | Use fileTitle to identify which document the chunk belongs to |
80| Group related chunks | Same fileTitle = same document context |
81| Deprioritize mismatches | fileTitle unrelated to query AND score > 0.5 → rank lower |
82 
83## Ingestion
84 
85### ingest_file
86```
87ingest_file({ filePath: "/absolute/path/to/document.pdf" })
88```
89 
90### ingest_data
91```
92ingest_data({
93  content: "<html>...</html>",
94  metadata: { source: "https://example.com/page", format: "html" }
95})
96```
97 
98**Format selection** — match the data you have:
99- HTML string → `format: "html"`
100- Markdown string → `format: "markdown"`
101- Other → `format: "text"`
102 
103**Source format:**
104- Web page → Use URL: `https://example.com/page`
105- Other content → Use scheme: `{type}://{date}` or `{type}://{date}/{detail}`
106  - Examples: `clipboard://2024-12-30`, `chat://2024-12-30/project-discussion`
107 
108**HTML source options:**
109- Static page → LLM fetch
110- SPA/JS-rendered → Browser MCP
111- Auth required → Manual paste
112 
113Re-ingest same source to update. Use same source in `delete_file` to remove.
114 
115## References
116 
117For edge cases and examples:
118- [html-ingestion.md](references/html-ingestion.md) - URL normalization, SPA handling
119- [query-optimization.md](references/query-optimization.md) - Query patterns by intent
120- [result-refinement.md](references/result-refinement.md) - Contradiction resolution, chunking
121

Full transparency — inspect the skill content before installing.