convez.ai
MCP ServerBusca & PesquisaAvançado
mirusser

RFC RAG

mirusser/rfc-rag

Busca semântica em RFCs com precisão de seção e filtragem normativa

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server que indexa localmente ~9.800 RFCs com busca semântica (pgvector) e full-text, retornando seções precisas em vez de documentos inteiros. Combina embeddings de IA com filtragem de palavras-chave normativas RFC 2119 para identificar requisitos vinculantes em padrões de internet.

Instalação

$npx convez-ai mirusser/rfc-rag

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
71%
Último commit
há 1 mês
Versão
v1.1.0
Issues abertas
2
Contribuidores
2

Casos de uso

  • Auditorias de conformidade com padrões de internet (RFC)
  • Pesquisa técnica em especificações de protocolo sem ler documentos inteiros
  • Integração de conhecimento de RFCs em agentes de IA via MCP

Veredito

Ferramenta sólida para quem trabalha com padrões de internet. Combina busca semântica com análise de requisitos normativos de forma elegante, mas requer setup de PostgreSQL+pgvector e indexação inicial (~15 min). Muito útil para compliance e pesquisa técnica profunda.

README

A tradução para o português deste README ainda está na fila.

📡 RFC RAG

Local RAG MCP server for RFCs: semantic search with section-level precision.

Ask RFCs, get answers:

Semantic search finds relevant RFCs.
Full-text search catches exact keywords.
Section-level retrieval returns the precise paragraph, not the whole document.
AI agents cite RFCs without copying 200 pages of spec.

CI License

.NET 10 PostgreSQL Docker MCP OpenRouter

📝 TL;DR

Index ~9,800 RFCs locally with pgvector + PostgreSQL full-text search, then query them from Claude Code or Codex (or any MCP compatible client) via MCP tools.

Semantic search understands meaning, full-text catches exact terms, and hybrid retrieval (RRF) combines both. Section-level indexing means you get the relevant paragraph, not a 200-page PDF.

Indexing takes ~10–15 minutes on first run. Incremental runs skip already-indexed files (SHA256-based) and complete in seconds.

Why?

RFCs are the backbone of internet standards, but finding the right section in the right RFC is painful. CTRL+F across 9,800 text files works poorly when you don't know the exact term. Semantic search understands that "how to structure JSON Web Tokens" means RFC 7519, even if the word "structure" never appears.

The real gap is in compliance auditing. Say you're a security engineer who needs to find every RFC section about encryption that prohibits something. You have three bad options:

  1. Read all ~9,800 RFCs — impossible.
  2. CTRL+F for "MUST NOT" — 682,664 results, most about unrelated topics.
  3. Semantic search for "encryption prohibition" — finds relevant sections, but many don't actually contain a formal prohibition. You're still guessing which are binding requirements vs. casual discussion.

This repo combines both: semantic search finds the topic, normative keyword filtering keeps only sections with RFC 2119/8174 requirement-level keywords (MUST, SHOULD, MUST NOT, etc.). The result is precise, citeable sections — you know exactly which RFC, which section, and which keyword makes it a binding requirement.

🗺️ Architecture

---
title: RFC RAG — Indexing and Query Flow
---
flowchart TB
    subgraph source["📁 RFC Mirror (~/OtherRepos/rfc-mirror/)"]
        Txt["📄 *.txt files\n~9,800 RFCs"]
    end

    subgraph index["🔍 Indexing Pipeline"]
        Parser["🔧 RfcParser\nsection splitter · metadata · ABNF · normative keywords"]
        Embed["🧠 Embedding Generator\nOpenRouter · text-embedding-3-small"]
    end

    subgraph store["🗄️ PostgreSQL + pgvector"]
        Sections["rfc_sections\nvector(1536) + tsvector"]
        Norm["normative_occurrences\nkeyword index"]
        Abnf["rfc_abnf_blocks\ngrammar search"]
    end

    subgraph search["🔎 Hybrid Search"]
        Vector["Cosine similarity\n(vector)"]
        FTS["Full-text search\n(tsvector)"]
        RRF["Reciprocal Rank Fusion"]
    end

    subgraph serve["📡 MCP Server"]
        Tools["🔧 MCP Tools"]
    end

    Clients["🤖 AI Agents"]

    Txt --> Parser
    Parser -->|"section text"| Embed -->|"(1536,) vector"| Sections
    Parser -->|"metadata · ABNF · keywords"| Sections
    Parser --> Norm
    Parser --> Abnf
    Sections --> Vector
    Sections --> FTS
    Vector --> RRF
    FTS --> RRF
    RRF --> Tools
    Norm --> Tools
    Abnf --> Tools
    Tools <-->|"MCP stdio"| Clients

The parser extracts sections, metadata, normative keywords, and ABNF grammar blocks from RFC text files. Section text is embedded via OpenRouter (text-embedding-3-small, 1536-dim) and stored alongside tsvector for full-text search. Hybrid search fuses vector cosine similarity with lexical full-text scores using Reciprocal Rank Fusion (RRF). A separate MCP stdio server exposes 10 tools for AI agents.

⚡ Quick Start

🐳 Docker Compose

cp deploy/compose/rfc-rag.env.example .env.rfc-rag
# edit .env.rfc-rag
make quickstart

Or directly:

docker compose --env-file .env.rfc-rag -f deploy/compose/rfc-rag.yaml up

Full configuration guide: configuration.md

🧰 MCP Tools

🔎 Search & Retrieval

ToolPurpose
search_rfcHybrid search (vector + full-text) with RRF fusion. Supports normative_keyword filtering.
get_rfcRFC metadata, table of contents, and section preview
get_rfc_sectionSpecific section with child expansion for nested subsections
get_rfc_tocTable of contents as section → heading map
get_rfc_metadataSingle RFC metadata lookup (title, authors, date, status)

📊 Analysis

ToolPurpose
search_normativeSearch normative keywords (MUST, SHOULD, MUST NOT, SHALL, etc.) across all RFCs. See docs/normative-search.md for how normative keyword extraction and filtering work under the hood.
search_abnfSearch extracted ABNF grammar definitions
find_updates_obsoletesBack-reference lookup — find RFCs that update or obsolete a given RFC
rfc_statsIndexed corpus statistics (total RFCs, sections, keywords, embeddings)
list_indexed_rfcsPaginated list of indexed RFCs with metadata

Full tool documentation in src/RfcRag/README.md.

💻 CLI Mode

Pass --cli <verb> [args] to run a one-shot query against an indexed database instead of starting the MCP server. All output is JSON on stdout.

Full CLI mode guide: cli-mode-guide.md

⌨️ Connecting AI Agents

Claude Code

claude mcp add-json --scope user rfc-rag \
  '{"type":"stdio","command":"dotnet","args":["run","--project","src/RfcRag/"]}'

Codex

# ~/.codex/config.toml
[mcp_servers.rfc-rag]
command = "dotnet"
args = ["run", "--project", "src/RfcRag/"]

Containerized MCP

[!NOTE] Run the server in Docker, then connect via docker exec.

claude mcp add-json --scope user rfc-rag \
  '{"type":"stdio","command":"docker","args":["exec","-i","rfc-rag-rfc-rag-1","dotnet","RfcRag.dll"]}'

🗄️ Database Schema

rfc_rag.rfc_sections           — primary search unit (vectors + FTS)
rfc_rag.indexed_rfcs           — SHA256 tracking for incremental indexing
rfc_rag.rfc_abnf_blocks        — extracted ABNF grammar blocks
rfc_rag.normative_occurrences  — pre-extracted normative keywords
rfc_rag.schema_migrations      — applied migration tracking

🧪 Running Tests

# All tests (unit + integration)
make test

# Unit tests only (no dependencies)
dotnet test --filter "Category!=Integration"

# Integration tests (requires Docker)
dotnet test --filter "Category=Integration"

# Retrieval quality suite — indexes TestData corpus and asserts top-10 hit rate (requires Docker)
dotnet test tests/RfcRag.Tests/ --filter "Category=RetrievalQuality"

🧩 Compatibility

AreaSupported / tested
.NET.NET 10
PostgreSQL15+ with pgvector 0.5+
MCP transportstdio (ModelContextProtocol 1.3.0)
EmbeddingsOpenRouter (text-embedding-3-small, 1536-dim)
Platformslinux/amd64, linux/arm64
DockerCompose v2, standalone

🧭 Project Map

⚖️ Boundaries

[!IMPORTANT]

  • Indexes a local RFC mirror — does not fetch RFCs from the internet at query time.
  • Embeddings are generated via OpenRouter by default (requires internet during indexing); set RfcRag__EmbeddingProvider=Local to use a local server (Ollama, llama.cpp) with no internet required.
  • MCP transport is stdio-only — no HTTP endpoint exposed.
  • The RAG pipeline answers from indexed RFC content; it does not perform live web search or access external knowledge bases.
  • This is a local development and research tool, not a production-certified service.

📜 Governance


Ver o original no GitHub

Bom para

Relacionados