PAMPA – Protocol for Augmented Memory of Project Artifacts

Version 1.12.x · Semantic Search · MCP Compatible · Node.js

Give your AI agents an always-updated, queryable memory of any codebase – with intelligent semantic search and automatic learning – in one npx command.

🇪🇸 Versión en Español | 🇺🇸 English Version | 🤖 Agent Version

🌟 What's New in v1.12 - Advanced Search & Multi-Project Support

🎯 Scoped Search Filters - Filter by path_glob, tags, lang for precise results

🔄 Hybrid Search - BM25 + Vector fusion with reciprocal rank blending (enabled by default)

🧠 Cross-Encoder Re-Ranker - Transformers.js reranker for precision boosts

👀 File Watcher - Real-time incremental indexing with Merkle-like hashing

📦 Context Packs - Reusable search scopes with CLI + MCP integration

🛠️ Multi-Project CLI - --project and --directory aliases for clarity

🏆 Performance Analysis - Architectural comparison with general-purpose IDE tools

Major improvements:

• 40% faster indexing with incremental updates

• 60% better precision with hybrid search + reranker

• 3x faster multi-project operations with explicit paths

• 90% reduction in duplicate function creation with symbol boost

• Specialized architecture for semantic code search

🌟 Why PAMPA?

Large language model agents can read thousands of tokens, but projects easily reach millions of characters. Without an intelligent retrieval layer, agents:

• Recreate functions that already exist

• Misname APIs (newUser vs. createUser)

• Waste tokens loading repetitive code (vendor/, node_modules/...)

• Fail when the repository grows

PAMPA solves this by turning your repository into a semantic code memory graph:

1. Chunking – Each function/class becomes an atomic chunk

2. Semantic Tagging – Automatic extraction of semantic tags from code context

3. Embedding – Enhanced chunks are vectorized with advanced embedding models

4. Learning – System learns from successful searches and caches intentions

5. Indexing – Vectors + semantic metadata live in local SQLite

6. Codemap – A lightweight pampa.codemap.json commits to git so context follows the repo

7. Serving – An MCP server exposes intelligent search and retrieval tools

Any MCP-compatible agent (Cursor, Claude, etc.) can now search with natural language, get instant responses for learned patterns, and stay synchronized – without scanning the entire tree.

🤖 For AI Agents & Humans

🤖 If you're an AI agent: Read the complete setup guide for agents → or 👤 If you're human: Share the agent setup guide with your AI assistant to automatically configure PAMPA!

📚 Table of Contents

• 🚀 MCP Installation (Recommended)

• 🧠 Semantic Features

• 📝 Supported Languages

• 💻 Direct CLI Usage

• 🧠 Embedding Providers

• 🏆 Performance Benchmark

• 🏗️ Architecture

• 🔧 Available MCP Tools

• 📊 Available MCP Resources

• 🎯 Available MCP Prompts

🧠 Semantic Features

🏷️ Automatic Semantic Tagging

PAMPA automatically extracts semantic tags from your code without any special comments:

Automatic tags: ["stripe", "service", "payment", "checkout", "session", "create"]

🎯 Intention-Based Direct Search

The system learns from successful searches and provides instant responses:

📈 Adaptive Learning System

• Automatic Learning: Saves successful searches (>80% similarity) as intentions

• Query Normalization: Understands variations: "create" = "crear", "session" = "sesion"

• Pattern Recognition: Groups similar queries: "[PROVIDER] payment session"

🏷️ Optional @pampa-comments (Complementary)

Enhance search precision with optional JSDoc-style comments:

Benefits:

• +21% better precision when present

• Perfect scores (1.0) when query matches intent exactly

• Fully optional: Code without comments works automatically

• Retrocompatible: Existing codebases work without changes

📊 Search Performance Results

📝 Supported Languages

PAMPA can index and search code in several languages out of the box:

• JavaScript / TypeScript (.js, .ts, .tsx, .jsx)

• PHP (.php)

• Python (.py)

• Go (.go)

• Java (.java)

🚀 MCP Installation (Recommended)

1. Configure your MCP client

Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

Optional: Add "--debug" to args for detailed logging: ["-y", "pampa", "mcp", "--debug"]

Cursor

Configure Cursor by creating or editing the mcp.json file in your configuration directory:

2. Let your AI agent handle the indexing

Your AI agent should automatically:

• Check if the project is indexed with get_project_stats

• Index the project with index_project if needed

• Keep it updated with update_project after changes

Need to index manually? See Direct CLI Usage section.

3. Install the usage rule for your agent

Additionally, install this rule in your application so it uses PAMPA effectively:

Copy the content from RULE_FOR_PAMPA_MCP.md into your agent or AI system instructions.

4. Ready! Your agent can now search code

Once configured, your AI agent can:

💻 Direct CLI Usage

For direct terminal usage or manual project indexing:

Install the CLI

Index or update a project

Indexing writes .pampa/ (SQLite database + chunk store) and pampa.codemap.json. Commit the codemap to git so teammates and CI re-use the same metadata.

| Command | Purpose | | ---------------------------------------- | -------------------------------------------------------- | ----- | ------------------------------------------------- | | npx pampa index [path] [--provider X] | Create or refresh the full index at the provided path | | npx pampa update [path] [--provider X] | Force a full re-scan (helpful after large refactors) | | npx pampa watch [path] [--provider X] | Incrementally update the index as files change | | npx pampa search <query> | Hybrid BM25 + vector search with optional scoped filters | | npx pampa context <list | show | use> | Manage reusable context packs for search defaults | | npx pampa mcp | Start the MCP stdio server for editor/agent integrations |

Search with scoped filters & ranking flags

pampa search supports the same filters used by MCP clients. Combine glob patterns, semantic tags, language filters, provider overrides, and ranking controls:

| Flag / option | Effect | | --------------------- | --------------------------------------------------------------------- | --------------- | | --path_glob | Limit results to matching files ("app/Services/**") | | --tags | Filter by codemap tags (stripe, checkout) | | --lang | Filter by language (php, ts, py) | | --provider | Override embedding provider for the query (openai, transformers) | | --reranker | Reorder top results with the Transformers cross-encoder (off | transformers) | | --hybrid / --bm25 | Toggle reciprocal-rank fusion or the BM25 candidate stage (on | off) | | --symbol_boost | Toggle symbol-aware ranking boost that favors signature matches (on | off) | | -k, --limit | Cap returned results (defaults to 10) |

PAMPA extracts function signatures and lightweight call graphs with tree-sitter. When symbol boosts are enabled, queries that mention a specific method, class, or a directly connected helper will receive an extra scoring bump.

When a context pack is active, the CLI prints the pack name before executing the search. Any explicit flag overrides the pack defaults.

Manage context packs

Store JSON packs in .pampa/contextpacks/*.json to capture reusable defaults:

MCP tip: The MCP tool use_context_pack mirrors the CLI. Agents can switch packs mid-session and every subsequent search_code call inherits those defaults until cleared.

Watch and incrementally re-index

The watcher batches filesystem events, reuses the Merkle hash store in .pampa/merkle.json, and only re-embeds touched files. Press Ctrl+C to stop.

Run the synthetic benchmark harness

The harness seeds a deterministic Laravel + TypeScript corpus and prints a summary table with Precision@1, MRR@5, and nDCG@10 for Base, Hybrid, and Hybrid+Cross-Encoder modes. Customise scenarios via flags or environment variables:

• npm run bench -- --hybrid=off – run vector-only evaluation

• npm run bench -- --reranker=transformers – force the cross-encoder

• PAMPA_BENCH_MODES=base,hybrid npm run bench – limit to specific modes

• PAMPA_BENCH_BM25=off npm run bench – disable BM25 candidate generation

Benchmark runs never download external models when PAMPA_MOCK_RERANKER_TESTS=1 (enabled by default inside the harness).

An end-to-end context pack example lives in examples/contextpacks/stripe-backend.json.

🧠 Embedding Providers

PAMPA supports multiple providers for generating code embeddings:

Recommendation: Use Transformers.js for personal development (free and private) or OpenAI for maximum quality.

🏆 Performance Analysis

PAMPA v1.12 uses a specialized architecture for semantic code search with measurable results.

📊 Performance Metrics

Synthetic Benchmark Results:

🎯 Search Examples

📈 Read Full Analysis →

🚀 Architectural Advantages

1. Specialized Indexing - Persistent index with function-level granularity

2. Hybrid Search - BM25 + Vector + Cross-encoder reranking combination

3. Code Awareness - Symbol boosting, AST analysis, function signatures

4. Multi-Project - Native support for context across different codebases

Result: Optimized architecture for semantic code search with verifiable metrics.

🏗️ Architecture

Key Components

🔧 Available MCP Tools

The MCP server exposes these tools that agents can use:

search_code

Search code semantically in the indexed project.

• Parameters:

  • query (string) - Semantic search query (e.g., "authentication function", "error handling")

  • limit (number, optional) - Maximum number of results to return (default: 10)

  • provider (string, optional) - Embedding provider (default: "auto")

  • path (string, optional) - PROJECT ROOT directory path where PAMPA database is located

• Database Location: {path}/.pampa/pampa.db

• Returns: List of matching code chunks with similarity scores and SHAs

get_code_chunk

Get complete code of a specific chunk.

• Parameters:

  • sha (string) - SHA of the code chunk to retrieve (obtained from search_code results)

  • path (string, optional) - PROJECT ROOT directory path (same as used in search_code)

• Chunk Location: {path}/.pampa/chunks/{sha}.gz or {sha}.gz.enc

• Returns: Complete source code

index_project

Index a project from the agent.

• Parameters:

  • path (string, optional) - PROJECT ROOT directory path to index (will create .pampa/ subdirectory here)

  • provider (string, optional) - Embedding provider (default: "auto")

• Creates:

  • {path}/.pampa/pampa.db (SQLite database with embeddings)

  • {path}/.pampa/chunks/ (compressed code chunks)

  • {path}/pampa.codemap.json (lightweight index for version control)

• Effect: Updates database and codemap

update_project

🔄 CRITICAL: Use this tool frequently to keep your AI memory current!

Update project index after code changes (recommended workflow tool).

• Parameters:

  • path (string, optional) - PROJECT ROOT directory path to update (same as used in index_project)

  • provider (string, optional) - Embedding provider (default: "auto")

• Updates:

  • Re-scans all files for changes

  • Updates embeddings for modified functions

  • Removes deleted functions from database

  • Adds new functions to database

• When to use:

  • ✅ At the start of development sessions

  • ✅ After creating new functions

  • ✅ After modifying existing functions

  • ✅ After deleting functions

  • ✅ Before major code analysis tasks

  • ✅ After refactoring code

• Effect: Keeps your AI agent's code memory synchronized with current state

get_project_stats

Get indexed project statistics.

• Parameters:

  • path (string, optional) - PROJECT ROOT directory path where PAMPA database is located

• Database Location: {path}/.pampa/pampa.db

• Returns: Statistics by language and file

📊 Available MCP Resources

pampa://codemap

Access to the complete project code map.

pampa://overview

Summary of the project's main functions.

🎯 Available MCP Prompts

analyze_code

Template for analyzing found code with specific focus.

find_similar_functions

Template for finding existing similar functions.

🔍 How Retrieval Works

• Vector search – Cosine similarity with advanced high-dimensional embeddings

• Summary fallback – If an agent sends an empty query, PAMPA returns top-level summaries so the agent understands the territory

• Chunk granularity – Default = function/method/class. Adjustable per language

📝 Design Decisions

• Node only → Devs run everything via npx, no Python, no Docker

• SQLite over HelixDB → One local database for vectors and relations, no external dependencies

• Committed codemap → Context travels with repo → cloning works offline

• Chunk granularity → Default = function/method/class. Adjustable per language

• Read-only by default → Server only exposes read methods. Writing is done via CLI

🧩 Extending PAMPA

🔐 Encrypting the Chunk Store

PAMPA can encrypt chunk bodies at rest using AES-256-GCM. Configure it like this:

1. Export a 32-byte key in base64 or hex form:

   export PAMPA_ENCRYPTION_KEY="$(openssl rand -base64 32)"

2. Index with encryption enabled (skips plaintext writes even if stale files exist):

   npx pampa index --encrypt on

   Without --encrypt, PAMPA auto-encrypts when the environment key is present. Use --encrypt off to force plaintext (e.g., for debugging).

3. All new chunks are stored as .gz.enc and require the same key for CLI or MCP chunk retrieval. Missing or corrupt keys surface clear errors instead of leaking data.

Existing plaintext archives remain readable, so you can enable encryption incrementally or rotate keys by re-indexing.

🤝 Contributing

1. Fork → create feature branch (feat/...)

2. Run npm test (coming soon) & npx pampa index before PR

3. Open PR with context: why + screenshots/logs

All discussions on GitHub Issues.

📜 License

MIT – do whatever you want, just keep the copyright.

Happy hacking! 💙

🇦🇷 Made with ❤️ in Argentina | 🇦🇷 Hecho con ❤️ en Argentina