HEngram: High-Performance Knowledge Graph MCP Server

HEngram is a specialized Model Context Protocol (MCP) server that provides a high-performance, persistent knowledge graph with hybrid search capabilities. Built for maximum efficiency with a low hardware footprint, it leverages a unique dual-process architecture combining Haskell for protocol safety and C++ for bare-metal embedding inference.

🚀 Key Features

• Hybrid Search Engine: Combines BM25 text matching and KNN vector similarity in a single Redis query for maximum recall and precision.
• Server-Side Intelligence: Built-in logic for semantic deduplication, automatic search fallback, and relationship validation.
• Managed Llama Supervisor: Automatically manages a local llama-server (C++) instance for generating high-quality embeddings via an OpenAI-compatible API.
• Matryoshka Truncation (MRL): Uses the nomic-embed-text-v1.5 model, truncated to 384 dimensions. This provides a significant speed boost and storage reduction with minimal accuracy loss.
• Graph Relationships: Explicit support for linking nodes with typed edges, allowing LLMs to build and navigate a structured knowledge network.
• Pure Functional Core: Written in Haskell for high concurrency, type-safe JSON-RPC handling, and predictable performance.

🧠 Server-Side Intelligence

HEngram moves critical graph logic from the LLM to the server to ensure data integrity and performance:

• Semantic Deduplication: When using memorize, HEngram checks for existing nodes with high semantic similarity within the same domain. If a duplicate is detected, it returns the existing node ID instead of creating a new one.
• Automatic Search Fallback: If a domain-specific search yields no results, HEngram automatically expands the query to a global search across all domains.
• Ghost Edge Prevention: The link tool validates that both the source and target nodes exist before creating a relationship, preventing "dangling" or "ghost" edges in the graph.

🏗️ Architecture

graph TD
    Client[MCP Client/Witsy] <-->|SSE/JSON-RPC| Servant[Haskell Servant Server]
    Servant <-->|Supervisor| Llama[llama-server C++]
    Servant <-->|Raw RESP| RediSearch[Redis / RediSearch]
    Llama <-->|Weights| Model[nomic-embed-text GGUF]
    RediSearch <-->|BM25 + KNN| Memory[Knowledge Graph]

Loading

Why Haskell + C++?

Many embedding systems rely on heavy Python stacks. HEngram uses Haskell's lightweight threads (Green Threads) to handle concurrent MCP sessions while delegating the heavy math to llama.cpp—the gold standard for CPU-based inference. This keeps the memory footprint low while maximizing performance across diverse environments.

🛠️ Tools Provided

HEngram exposes five primary tools to the LLM:

1. memorize: Stores information into the graph with automatic deduplication.
   • Inputs: domain, content, node_type (all required).
2. search: Performs a hybrid search with automatic global fallback.
   • Inputs: query (required), domain (optional).
3. link: Connects two existing nodes with existence validation.
   • Inputs: source_id, target_id, rel_type (all required).
4. recall: Retrieves a specific node by ID, including its full content and outgoing relationships.
   • Inputs: node_id (required).
5. patch: Updates fields of an existing node. Only specified fields are changed.
   • Inputs: node_id (required), content, domain, node_type (optional).

🏷️ Canonical Taxonomy

HEngram enforces a canonical taxonomy to keep the graph structured and searchable. Input values are automatically normalized to the closest match:

• Domains: personal, swe, infra, health, finance, recon, fitness, research, misc.
• Node Types: fact, preference, project, person, decision, constraint, event, concept, procedure, summary.

⚙️ Setup

Prerequisites

• GHC 9.6.6+ & Cabal
• Redis with RediSearch module enabled (Default port: 6379)
• curl & unzip (for automated binary setup)

Installation

1. Clone the repository:

   git clone https://github.com/di5rupt0r/HEngram.git
   cd HEngram

2. Build the project:

   cabal build

3. Run the server:

   cabal run hengram

   Note: On the first run, HEngram will automatically download llama-server and the embedding model. The embedding server runs on port 8081 by default.

📄 License

This project is licensed under the MIT License.