Skip to main content
Glama
omni-rag-mcp

omni-rag-mcp

by Suyash2013

omni-rag-mcp

A general-purpose RAG MCP plugin for token-efficient semantic search over any directory of files. Auto-ingests the current working directory on first search and provides hybrid search (BM25 + semantic), directory overview, structural analysis, and dependency graphs.

Zero-config by default: local Qdrant storage, ONNX embeddings, no external services required. Supports code, markdown, PDFs, CSVs, and more via pluggable extractors.

Quick Start

pip install omni-rag-mcp
omni-rag-setup

That's it. Restart Claude Code and the plugin auto-indexes your working directory on first search.

Related MCP server: Paparats: Local Semantic Code Search

How It Works

Your Files  ->  Extractors  ->  Chunking  ->  Embedding  ->  Qdrant (local)
                                                                 |
Claude Code ->  MCP Tool Call  ->  Hybrid Search  ->  Relevant Snippets
  1. First search auto-ingests your working directory (extracts content, chunks, generates embeddings, stores in local Qdrant)

  2. Subsequent searches are fast hybrid lookups (BM25 + semantic) -- no re-ingestion needed

  3. Incremental updates detect git changes and only re-embed modified files

MCP Tools

Tool

Purpose

search

Hybrid search over indexed files (auto-ingests if needed)

search_by_file

Search filtered by file path pattern

get_context

Compressed directory overview (languages, structure, dependencies)

get_file_signatures

Function/class signatures without reading every file

get_dependency_graph

Internal import/dependency graph

stats

Index size and configuration

ingest

Manual re-index (incremental by default, force=True for full)

check_status

Is the index current? Any uncommitted changes?

Embedding Providers

Zero-config by default. Choose your provider:

Provider

Config

Notes

ONNX (default)

None needed

Auto-downloads all-MiniLM-L6-v2 (23MB, 384-dim)

Ollama

OMNI_RAG_EMBEDDING_PROVIDER=ollama

Requires Ollama running with model pulled

OpenAI

OMNI_RAG_EMBEDDING_PROVIDER=openai + OMNI_RAG_OPENAI_API_KEY=sk-...

text-embedding-3-small

Voyage

OMNI_RAG_EMBEDDING_PROVIDER=voyage + OMNI_RAG_VOYAGE_API_KEY=...

voyage-code-3 (optimized for code)

Optional Extras

pip install omni-rag-mcp[pdf]    # PDF extraction (PyMuPDF)
pip install omni-rag-mcp[docx]   # Word document extraction
pip install omni-rag-mcp[image]  # Image/OCR extraction (Tesseract + Pillow)
pip install omni-rag-mcp[all]    # All optional extractors

Storage

By default, uses Qdrant in local/on-disk mode -- no Docker needed. Data stored in .omni-rag/ under your project directory.

For remote Qdrant:

OMNI_RAG_QDRANT_MODE=remote
OMNI_RAG_QDRANT_HOST=your-host
OMNI_RAG_QDRANT_PORT=6333

Configuration

All settings via environment variables with OMNI_RAG_ prefix. See config/.env.example for the full reference.

Legacy RAG_ prefix variables are still supported with deprecation warnings.

Development

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
python -m pytest tests/ -v

# Health check
python scripts/health_check.py

Manual MCP Registration

If omni-rag-setup doesn't work, add this to your Claude Code MCP config:

{
  "mcpServers": {
    "omni-rag": {
      "command": "omni-rag"
    }
  }
}

Available Tools

14 tools
check_index_statusA

[Deprecated: use 'check_status'] Check index freshness.

Inputs: no named parameters. Returns: result

check_statusA

Check whether the index for the current directory is up to date.

Call this FIRST if you're unsure whether to use semantic search or fall back to file-by-file reading. Returns whether the directory is indexed and whether there are local changes that might make the index stale.

Inputs: no named parameters. Returns: result

collection_statsA

[Deprecated: use 'stats'] Get index statistics.

Inputs: no named parameters. Returns: result

get_codebase_contextA

[Deprecated: use 'get_context'] Get directory overview.

Inputs: no named parameters. Returns: result

get_contextA

Get a compressed overview of the indexed directory structure.

Returns language breakdown, directory structure, key files, and dependency information. Use this FIRST when starting work on an unfamiliar directory to understand its layout before searching for specific content.

Much cheaper than reading files individually — gives you the big picture in one call.

Inputs: no named parameters. Returns: result

get_dependency_graphA

Get the import/dependency graph for the codebase.

Shows which files import from which other files within the project. Use this to understand module boundaries and find the right place to make changes.

Args: file_pattern: Optional substring to filter files (case-insensitive). Empty string analyzes the entire project.

Inputs: file_pattern. Returns: result

get_file_signaturesA

Get function and class signatures from files matching a path.

Use this to understand the API surface of a module without reading every file. Returns function names, parameters, return types, and class hierarchies.

More structured than semantic search — gives you the exact interface of code modules.

Args: file_path: Substring to match in file paths (case-insensitive). E.g. "models", "api/routes", ".py", "utils". Empty string matches all files.

Inputs: file_path. Returns: result

ingestA

Manually trigger indexing of the current working directory.

Auto-indexes on first search, so you rarely need this. By default, does incremental indexing (only changed files). Use force=True to do a full re-index from scratch.

Args: force: If True, full re-index. If False, incremental update. include_extensions: Optional list of file extensions to include (e.g. [".py", ".ts", "js"]). When provided, ONLY files with these extensions are indexed; the default TEXT_EXTENSIONS allowlist is replaced. Leading dots are optional. exclude_extensions: Optional list of file extensions to exclude (e.g. [".txt", ".md"]). Files with these extensions are skipped even if they would otherwise be indexed. Leading dots are optional.

Inputs: force, include_extensions, exclude_extensions. Returns: result

ingest_current_directoryB

[Deprecated: use 'ingest'] Manually trigger indexing.

Inputs: force, include_extensions, exclude_extensions. Returns: result

search_by_fileA

Search indexed files, restricting results to paths matching a pattern.

Args: query: Natural language query. file_pattern: Glob-style pattern or substring for file paths (e.g. '.py', 'models/'). n_results: Optional number of results to return (default: 10).

Inputs: query (required), file_pattern (required), n_results. Returns: result

search_codebaseB

[Deprecated: use 'search'] Semantic search over indexed files.

Inputs: query (required), n_results. Returns: result

search_codebase_by_fileB

[Deprecated: use 'search_by_file'] Search filtered by file pattern.

Inputs: query (required), file_pattern (required), n_results. Returns: result

statsA

Get statistics about the current index.

Returns chunk count, collection name, storage mode, and embedding provider. Use this to verify the index is loaded and check its size.

Inputs: no named parameters. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.4/5.0

    Scored across 14 tools

    Disambiguation2/5

    Several tools are exact duplicates of others, with one marked deprecated and one current (e.g., check_index_status vs check_status). This creates ambiguity as an agent cannot easily distinguish which to use, leading to potential misselection.

    Naming Consistency3/5

    Naming is somewhat consistent with a verb_noun pattern (e.g., search, search_by_file, ingest, stats), but deprecated tools use inconsistent forms like check_index_status and get_codebase_context, breaking the pattern.

    Tool Count4/5

    With 14 tools, the count is reasonable for a RAG server. However, 6 of them are deprecated, reducing the effective set to 8 well-scoped tools, which is appropriate for the domain.

    Completeness4/5

    The toolset covers key operations: index status, context overview, search (including file-filtered), statistics, ingestion, dependency graph, and file signatures. Missing explicit index deletion or update, but core workflows are covered.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Provides LLM-optimized tools for advanced code analysis, repository complexity evaluation, and call graph generation. It enables users to visualize directory structures, detect code patterns, and build semantic context with significant token savings.
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Self-hosted semantic code search for AI coding agents. Indexes many repositories at once with tree-sitter AST chunking, local llama.cpp embeddings and Qdrant, so Claude Code, Cursor, Codex and any MCP client can find code by meaning, trace usages and callers, and read the git history behind each chunk. No API keys, and your code never leaves your machine.
      2
      18
      34
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables agents to perform local hybrid code search and code intelligence across a workspace, including semantic and full-text search, symbol lookup, file outlines, and caller analysis.
      MIT
    • A
      license
      Not graded
      quality
      F
      maintenance
      Enables AI agents to retrieve precise structural and semantic context from codebases using hybrid search across 166+ languages. Supports offline/local operation and automatic fallback for resilient, token-efficient code understanding.
      35 PyPI
      12
      Apache 2.0