Skip to content

RAG & Memory Subsystem

This feature covers conversation memory management, vector search with pgvector, document chunking, and MinIO S3 object storage.

1. Dual-Layer Memory Architecture

graph TD
    subgraph Short-Term Memory
        STM[session_short_memory<br/>Recent conversation history]
    end

    subgraph Long-Term Vector Memory
        LTM[session_long_memory / rag_chunks<br/>Pgvector 1536d Cosine Similarity]
    end

    subgraph Object Storage
        S3[MinIO S3 / Local Storage<br/>Raw ingested document artifacts]
    end

    AgentExec[Agent Execution Pipeline] --> STM
    AgentExec --> LTM
    RagService[RagService Ingestion] --> S3
    RagService --> LTM

2. Conversation Memory (Short Memory)

  • Persisted per session in session_short_memory PostgreSQL table.
  • Stores role-based messages (user, assistant, system) with automatic sliding window truncation (shortMemorySize).

3. Vector Memory & RAG (Long Memory)

  • Document Chunking: RecursiveCharacterChunker splits documents into 1200-character segments with 150-character overlap across semantic boundaries (\n\n, \n, .).
  • Pgvector Indexing: Chunks and 1536-dimensional embeddings stored in rag_chunks table with HNSW cosine similarity index (vector_cosine_ops).
  • MinIO S3 Storage: Uploaded raw document binaries stored in MinIO S3 buckets (rag-documents) or local filesystem (LocalObjectStorage).