Knowledge Base Retrieval Tester in Sparkforge

By Sparkforge Support Team ·Last reviewed May 3, 2026

The Knowledge Base Retrieval Tester runs read-only test questions against a Sparkforge Knowledge Base and shows which sources or chunks are returned, so you can validate retrieval before connecting it to live AI agents.

The Knowledge Base Retrieval Tester runs sample questions against a Sparkforge Knowledge Base and shows which sources or chunks are retrieved. Use it to validate that documents, rich-text articles, tables, and crawled web content actually surface for the questions your customers will ask — before connecting the Knowledge Base to an AI Employee, Conversation AI bot, or Voice AI agent.

What the tester does

Real-time testing: ask a question and see how the Knowledge Base responds.
Source-level transparency: review the content sources or chunks used to generate the result.
Faster troubleshooting: spot missing, irrelevant, outdated, or poorly structured content before going live.
Read-only: tests don't change Knowledge Base content or affect live conversations.

When to use it

You added a new document, PDF, DOC, DOCX, rich-text article, CSV table, or website source.
You updated existing Knowledge Base content.
You're preparing an AI Employee, Conversation AI bot, or Voice AI agent for launch.
You noticed AI responses citing the wrong source or missing details.
You want to compare how different question phrasings retrieve different chunks.
You suspect duplicate or conflicting content is affecting retrieval.

Understanding retrieved sources and chunks

A Knowledge Base breaks source content into searchable pieces called chunks. Documents are parsed by headings and text. Tables are semantically indexed so plain-English questions can match rows.

When reviewing test results, check:

Whether the expected source appears.
Whether the returned chunk directly answers the question.
Whether irrelevant or outdated content is being retrieved.
Whether similar sources contain duplicate or conflicting information.
Whether content is too broad, too vague, or missing terms customers actually use.

Run a retrieval test

Navigate to AI Agents → Knowledge Base.
Open the Knowledge Base you want to test.
Confirm it includes the sources you want to validate — documents, rich-text, tables, web crawler content, or other supported types.
Wait for newly added or updated content to finish processing.
Open the Knowledge Base Retrieval Tester.
Enter a realistic customer-style question, like "What is your refund policy?" or "Which customers have overdue invoices?"
Review the answer and the retrieved sources.
If results are missing, incomplete, outdated, or irrelevant, update the Knowledge Base and retest the same question.

Best practices

Test simple, direct questions first.
Use natural customer wording, not internal terminology.
Ask the same question in multiple ways.
Include questions that should hit a specific document, table row, or policy.
Test edge cases and high-impact customer questions.
Review sources, not just the generated answer.
Remove or update outdated content that competes with newer entries.
Break long or unclear content into cleaner sections when retrieval is poor.
Retest after every major update.

Troubleshooting retrieval results

Issue	Likely cause	Action
No source retrieved	Content not indexed or phrased differently than the question.	Confirm indexing is complete; rephrase or add wording that matches user questions.
Wrong source retrieved	Another source has overlapping terms.	Update, remove, or clarify competing content.
Answer is incomplete	Source doesn't include all required details.	Add missing details to the relevant entry.
Outdated source retrieved	Old content still lives in the Knowledge Base.	Remove or update outdated entries.
Table data not returned	Table still processing or columns not indexed.	Confirm CSV setup, selected columns, and processing status.
Document content missing	Document wasn't parsed cleanly.	Review headings, formatting, and source quality; re-upload or revise.
Results improve after rewording	Knowledge Base lacks common customer phrasing.	Add natural-language terms, synonyms, and clearer headings.

Frequently Asked Questions

Does the tester change my Knowledge Base?

No. The tester is a read-only simulation. It doesn't edit, delete, or update Knowledge Base content.

Does testing affect live customer conversations?

No. Tests don't send messages to customers or trigger live conversations.

Is this the same as testing my Conversation AI bot?

No. Retrieval testing checks whether Knowledge Base content can be found and returned. Bot testing checks the broader conversation experience — goals, prompts, actions, and response behavior.

Why is the wrong source being retrieved?

Another source likely contains similar wording, duplicate information, outdated details, or broader content that appears more relevant. Review overlapping sources and make the correct one clearer.

What types of content can I test?

Documents, rich-text content, tables, and web-based sources — any supported Knowledge Base content type.

Does the tester show the exact answer customers will see?

It validates retrieval and source usage. The final customer-facing response also depends on the connected AI tool, prompt, bot setup, conversation context, and channel.