Why AI Agents Need Verifiable Evidence: Building an MCP-Native Retrieval Engine with PostgreSQL

Vinay Kumar K S — Fri, 18 Sep 2026 04:59:50 +0000

Why standard RAG fails agentic workflows, and how we built Knowledge Fabric using PostgreSQL, pgvector, and FastMCP.

Most Retrieval-Augmented Generation (RAG) setups treat context retrieval as a simple black box: prompt in, cosine similarity ranking, top-k text dump out.

In conversational chatbots, a slightly noisy chunk is usually harmless. But in agentic architectures—where models make tool calls, plan execution steps, and trigger real-world actions—retrieval without provenance is a major vulnerability. If an agent cannot verify where an excerpt originated, what version of a document it reflects, or whether the text was modified, policy checks fail and downstream execution becomes unreliable.

To address this, we built Knowledge Fabric: an open-source, vendor-neutral evidence retrieval engine designed natively for the Model Context Protocol (MCP) and built on PostgreSQL with pgvector.

The Problem: Naive RAG vs. Agentic Evidence

When an agent needs context to answer an enterprise query or decide on an action, standard vector stores present three common issues:

Exact-Term Blind Spots: Pure dense vector search often misses exact identifiers (e.g., invoice IDs, function signatures, error codes).
Missing Citations & Provenance: Agents receive text strings without cryptographic hashing, source revision IDs, or bounding offsets.
Bloated Context Windows: Shoveling sprawling context into prompts increases latency and token costs while confusing agent decision boundaries.

Architectural Design: The Three-Tier Fabric

Rather than building an all-in-one monolith that handles planning, retrieval, and write executions simultaneously, we separate concerns into three distinct roles:

Knowledge Fabric (The Librarian): Collects, indexes, searches, and provides verifiable evidence packages with clear citations.
Intent Fabric (The Planner & Referee): Translates user intent and evidence into policy-validated plans.
Enterprise Adapters (The Controlled Doors): Read from external systems and execute approved write mutations.

Workflow:

Storage Engine: Standard Infrastructure First

We deliberately avoided introducing custom daemons or proprietary vector databases. Knowledge Fabric is built directly on PostgreSQL + pgvector:

Hybrid Search with Reciprocal Rank Fusion (RRF): Fuses PostgreSQL BM25-style lexical search (tsvector) with semantic vector embeddings (pgvector) to ensure exact keyword recall alongside conceptual relevance.
Evidence Packaging: Every retrieved passage includes document URI, offset boundaries, chunk hash, and retrieval explanation metadata.
Native FastMCP Interface: Exposes retrieval directly over stdio and HTTP endpoints for Claude Code, Cursor, Codex, and custom agent harnesses.

FastMCP Tool Surface

Knowledge Fabric exposes five core MCP tools:

retrieve_evidence: Performs hybrid search and returns bounded, cited evidence packages.
get_document: Fetches full document content and metadata by ID or URI.
list_sources: Inspects currently indexed collections and document trees.
explain_retrieval: Returns the scoring breakdown (lexical vs. vector RRF ranks) for an evidence query.
health: Confirms database connectivity and vector index status.

3-Minute Quickstart

You can test Knowledge Fabric locally using Docker Compose:

Clone and launch PostgreSQL with pgvector:

git clone https://github.com/sagarv48/knowledge-fabric.git
cd knowledge-fabric
docker compose up -d

Install dependencies & initialize the database:

python -m venv .venv
source .venv/bin/activate
pip install -e .
python -m knowledge_fabric.cli init-db

Run the FastMCP Server: python -m knowledge_fabric.mcp.server

You can now add the server to your Claude Desktop or Cursor MCP configuration:

{
  "mcpServers": {
    "knowledge-fabric": {
      "command": "python",
      "args": ["-m", "knowledge_fabric.mcp.server"],
      "env": {
        "DATABASE_URL": "postgresql://postgres:postgres@localhost:5432/knowledge_fabric"
      }
    }
  }
}

What's Next & Getting Involved

Knowledge Fabric is early and actively developed. We are currently refining local embedding pipelines (Ollama integration) and chunking strategies.

GitHub Repository: https://github.com/sagarv48/knowledge-fabric
Roadmap & Issues: Check out our open issues for connector ideas and hybrid search tuning.

If you are building with MCP or experimenting with agent architectures, I would love to hear how you currently tackle citation and evidence verification in your workflows!

DEV Community: Vinay Kumar K S