KB Explorer¶

The KB Explorer is Beyond Retrieval's most advanced retrieval strategy. Instead of running a single search query, it deploys an AI agent that autonomously researches your question — browsing files, searching by keywords and meaning, reading documents, and reasoning about what it finds.

Think of it as an AI research assistant that explores your knowledge base the way a human would: look around, search, read, think, and repeat until it has enough information to give a thorough answer.

When to Use KB Explorer¶

KB Explorer shines when:

• Your question has multiple parts (e.g., "Compare Q3 and Q4 results and identify trends")
• The answer requires connecting information from different documents
• You're not sure which documents contain the answer
• The question is broad or exploratory

How It Works¶

When you send a message with KB Explorer enabled, the AI agent follows a research loop:

graph TD
    A["Your Question"] --> B["Break Down Question"]
    B --> C["Browse & Search"]
    C --> D["Read Documents"]
    D --> E{"Enough Info?"}
    E -->|"No — gaps found"| C
    E -->|"Yes — sufficient"| F["Generate Answer"]

1. Decompose — The agent breaks your question into smaller, focused sub-questions
2. Browse — It maps out what documents are available in your notebook
3. Search — It runs keyword and semantic searches for each sub-question
4. Read — It reads the most relevant document chunks
5. Reflect — It checks: "Do I have enough information to answer well?"
6. Iterate — If gaps exist, it searches again with refined queries
7. Answer — Once satisfied, it synthesizes everything into a comprehensive response

The 15 Tools¶

The KB Explorer agent has 15 specialized tools at its disposal. During a research session, you'll see these appear in the "Researching your question" panel.

Browse Tools¶

These tools help the agent understand what's in your knowledge base.

Search Tools¶

These tools find relevant content across your documents.

Read Tool¶

Reasoning Tools¶

These tools allow the agent to think, analyze, and evaluate what it has found.

Quality Check Tool¶

What You See in the Chat¶

When KB Explorer is active, you'll see a live progress panel showing each research step:

During Research (Live)¶

The panel shows each step as it happens:

• Step number and icon — Color-coded by category (blue for browse, amber for search, green for read, purple for reasoning)
• Friendly tool name — e.g., "Finding relevant passages" instead of technical names
• Description — What the tool is doing right now
• Result count — How many results were found (shown as a badge)
• Duration — How long each step took
• Sub-questions — When the agent breaks down your question, the sub-questions appear indented
• Confidence score — When the agent checks answer quality, a progress bar shows 0-100%

After Research (Collapsed)¶

Once the agent finishes researching, the panel collapses to a single line:

Researched 6 steps in 1.8s · Complete

Click it to expand and see the full research trail.

Typical Research Flows¶

Simple Question (3-5 steps)¶

1. Browsing documents     — Map the knowledge base
2. Searching for keywords — Find relevant content
3. Reading document       — Read the best match
→ Answer generated

Complex Question (8-12 steps)¶

1. Breaking down your question  — Split into 3 sub-questions
2. Browsing documents           — Map available files
3. Searching for keywords       — Search for sub-question 1
4. Finding relevant passages    — Semantic search for sub-question 2
5. Reading document             — Read Q3_Report.pdf, page 4
6. Reading document             — Read Annual_Review.pdf, page 12
7. Checking answer quality      — Score: 65% — gaps found
8. Expanding search terms       — Broaden search for missing info
9. Finding relevant passages    — New search with expanded query
10. Reading document            — Read Market_Analysis.docx
11. Checking answer quality     — Score: 92% — sufficient
→ Answer generated with citations from 3 sources

Configuration¶

Enable KB Explorer¶

In Notebook Settings > Intelligence > RAG Strategy, select KB Explorer.

Alternatively, via the API:

{
  "strategies_config": {
    "strategy_id": "kb-explorer"
  }
}

Parameters¶

How It Compares¶

Troubleshooting¶

Agent takes too long¶

Reduce max_steps to 8-10, or switch to a faster LLM model (e.g., openai/gpt-4o-mini).

Agent doesn't find relevant content¶

• Make sure your documents are ingested and embedded
• Try a different embedding model with better multilingual support
• Check if the document content was parsed correctly in the Documents page

"Checking answer quality" shows low score¶

This means the agent couldn't find enough relevant content. This usually indicates:

• The answer isn't in your documents
• The documents need to be re-ingested with a different parser
• Try rephrasing your question to be more specific