How It Works
At a high level, DeepBlock’s architecture has several interconnected components that work together to deliver fast, intelligent data access:
Data Ingestion DeepBlock operates blockchain indexers that connect to various networks (e.g. Ethereum, Layer-2s). These indexers stream blocks, transactions, and events into our system in near real-time. Raw data is parsed, cleaned, and normalized into a common format.
Knowledge Graph All indexed data is organized into a global knowledge graph of on-chain activity. In this graph, addresses, contracts, tokens, and other entities are represented as nodes, with relationships (like transfers, interactions, or ownership links) as edges. This semantic structure makes it easy to traverse connections and gather insights that would be hard to get from raw tables.
GraphQL Query Engine We expose the knowledge graph through a GraphQL API. Users can write rich queries that span multiple entities and even multiple chains in one request. The query engine translates GraphQL queries into efficient graph lookups. A performance layer (with caching and pre-computed aggregates) ensures that even complex analytical queries return quickly. You get exactly the data you ask for, with no need to manually aggregate or join across sources.
RAG Pipeline (Retrieval-Augmented Generation) For AI applications, DeepBlock integrates a RAG pipeline. This means when you ask a natural-language question or when an AI agent needs information, DeepBlock will retrieve the relevant on-chain data (using the knowledge graph and search indices) and feed it into a language model. The LLM (Large Language Model) then generates an answer or analysis based on current, factual data rather than just its training knowledge. This pipeline eliminates hallucinations by grounding the AI in up-to-date blockchain facts.
Model Context Protocol (MCP) To make the AI integration robust, DeepBlock uses a Model Context Protocol. MCP is essentially a structured format or set of guidelines for packaging the retrieved data and context before giving it to the LLM. It ensures that the LLM receives context in a consistent, easy-to-parse way. For example, clearly separating factual data (like numbers, addresses, relationships) from the question prompt. By enforcing this structure, MCP helps the AI reason correctly about the data and maintain awareness of on-chain relationships when formulating answers.
All these pieces work together seamlessly: data flows in from blockchains, gets structured into the knowledge graph, is queried via GraphQL or extracted via RAG for AI, with MCP ensuring the AI’s understanding.
The result is a conversational, intelligent data layer; you can either query it directly with GraphQL for programmatic use or ask questions and have an AI agent do the heavy lifting with reliable data behind it.
Last updated