• DATA
  • SEARCH
  • STRUCTURE
  • HUMAN

How to debug RAG retrieval with your AI agent

Connect Cognita to your agent to inspect the data being fed into your context window. You'll see the specific chunks that drive your model's answers.

Ask AI about this page

Short answer

Why is my AI pulling the wrong context?

You can use your agent to pull specific text chunks and check the underlying collections. This lets you see if the issue is bad data or a weak search query. You'll get a clear view of the retrieval pipeline.

Debugging outcomes.

Where every issue lands.

The agent identifies exactly where the retrieval process is breaking down.

DATA

Identified bad chunks

The AI finds specific text segments that don't match your intent. You'll know which parts of your documents are causing noise.

SEARCH

Query refinement

By seeing what the agent looks for, you can rewrite your prompts. You'll get better results by fixing how the agent searches.

STRUCTURE

Collection health check

The AI inspects how your data is organized. You'll see if your vector store is missing the context it needs to be useful.

HUMAN

Manual fixes needed

The AI flags where the source data is broken. You'll need to step in to clean up the documents or re-index the files.

The workflow

What your AI does when debugging starts.

The agent moves from high-level queries to granular data inspection.

  1. Run a test query

    The agent executes a search to see what the current system returns.

    rag_query

  2. Inspect the results

    The agent looks at the specific text segments that were pulled for the answer.

    search_chunks

  3. Verify the source

    The agent checks the specific group of documents being used for the search.

    get_collection

  4. Compare context

    The agent runs a new search to see if different parameters yield better chunks.

    rag_querysearch_chunks

Try it

Copy these to start.

Use these prompts to begin your debugging session.

Starting points

These are opening lines. Swap the bracketed text for your actual collection names or query terms.

Cognita Connector

You're all set. Choose your MCP client and follow the setup instructions.

Connector linkhttps://edge.vinkius.com/vk_preview_SHxG8KJRFQw2m3lX2GfeLS7zRF34WrsRFB0j95Nv/mcp

Claude Desktop

Follow the steps below to connect in seconds.

  1. 1In Claude Desktop, open Settings → Connectors.
  2. 2Click “Add custom connector” and paste the connector link above as the remote MCP server URL.
  3. 3Click Add and start a new chat — Cognita capabilities are ready to use.
Configuration · claude_desktop_config.jsonCopy
{
  "mcpServers": {
    "cognita-rag-framework-mcp": {
      "url": "https://edge.vinkius.com/vk_preview_SHxG8KJRFQw2m3lX2GfeLS7zRF34WrsRFB0j95Nv/mcp"
    }
  }
}
Copy into chat04

  • Run a rag_query for 'recent sales reports' and show me the chunks used.

  • Search the 'customer_feedback' collection for chunks related to 'latency'.

  • Get the details for the 'product_manual' collection and check its size.

  • Find the specific chunks in the 'legal_docs' collection that mention 'liability'.

  • Claude
  • ChatGPT
  • Cursor
  • VS Code
  • Windsurf
  • Claude Code
  • JetBrains
  • Cline

Start here

Connect Cognita once, then ask.

Just link your account once. Your credentials stay encrypted, and you can start debugging immediately in your favorite client.

Connect Cognita to your AI

FAQ

How this task behaves.

  • 01

    Can the AI change my document chunks?

    No. The AI can only read and inspect the data. You have to manually edit your documents or re-index them yourself.

  • 02

    Can the AI delete a collection?

    No. There are no tools provided to delete or modify your data collections.

  • 03

    Does this work with any vector database?

    It works with any data managed through the Cognita framework.

  • 04

    Can the AI fix a bad retrieval automatically?

    No. The AI identifies the problem, but you must decide how to fix your indexing or your prompts.

  • 05

    How do I see the raw text of a chunk?

    You can ask the agent to search for specific chunks to see the exact text it retrieved.

  • More questions about Cognita? The Connector page answers them. See everything the Cognita Connector can do

Connect Cognita to Claude, Cursor, ChatGPT & more