Context Graph Architecture

The Context Graph is the knowledge infrastructure backbone of the MATIH platform. It captures agent reasoning, entity relationships, and decision lineage in a multi-backend storage system that combines graph databases, vector stores, and bi-temporal relational storage.

High-Level Architecture

The Context Graph module resides within the AI Service and connects to multiple storage backends depending on the query type and data characteristics.

+---------------------------------------------------+
|              AI Service (Port 8000)                |
|                                                    |
|  +---------------------------------------------+  |
|  |           Context Graph Module               |  |
|  |                                              |  |
|  |  +------------------+ +-------------------+  |  |
|  |  | Agent Thinking   | | Orchestrator      |  |  |
|  |  | Service          | | Hooks             |  |  |
|  |  +--------+---------+ +--------+----------+  |  |
|  |           |                     |             |  |
|  |  +--------v---------+ +--------v----------+  |  |
|  |  | Dgraph Context   | | Kafka Producer /  |  |  |
|  |  | Store            | | Consumer          |  |  |
|  |  +--------+---------+ +--------+----------+  |  |
|  |           |                     |             |  |
|  |  +--------v---------+ +--------v----------+  |  |
|  |  | Hybrid Store     | | Metrics Bridge    |  |  |
|  |  | (GraphRAG)       | |                   |  |  |
|  |  +------------------+ +-------------------+  |  |
|  +---------------------------------------------+  |
+---------------------------------------------------+
        |              |              |
   +----v----+   +----v----+   +----v--------+
   | Dgraph  |   | Pinecone|   | PostgreSQL  |
   | (Graph) |   | (Vector)|   | (Bitemporal)|
   +---------+   +---------+   +-------------+

Storage Backends

The Context Graph uses three complementary storage backends, each optimized for different query patterns.

Multi-Tenant Isolation

Every operation in the Context Graph requires a tenant_id parameter. Isolation is enforced at each storage layer:

• Dgraph: GraphQL queries filter by tenantId field on all entity types
• Pinecone: Namespaces are prefixed with the tenant ID (format: tenant_id:namespace)
• PostgreSQL: All tables include a tenant_id column with indexed filters

Data Flow

Agent interactions flow through the Context Graph in a pipeline pattern:

1. Capture -- Orchestrator hooks intercept agent thinking steps and LLM calls
2. Persist -- The Thinking Service stores traces in Dgraph with optional Kafka streaming
3. Embed -- The Thinking Embedding Service generates vectors for input, output, and reasoning
4. Index -- Vectors are upserted into Pinecone namespaces for similarity search
5. Query -- The Semantic Search Service combines vector and graph results for retrieval

Key Components

GraphRAG Retrieval Strategies

The Hybrid Store implements multiple retrieval strategies based on the Microsoft GraphRAG architecture:

Feature Flag Control

Context Graph features are rolled out per-tenant using the Semantic Feature Flags service. Features can be toggled independently:

Rollout modes include disabled, canary (hash-based percentage), partial, and full.

Configuration

The Context Graph is configured through environment variables. No credentials are hardcoded.