Test and troubleshoot retrieval
Evaluate retrieval and answer generation separately. If an answer is wrong, first determine whether the correct source text was indexed and retrieved before changing the prompt or model.
Test in four layers
1. Source extraction
Under Indexed Sources, confirm every required source has completed processing and a nonzero chunk count. Open its chunks and verify names, dates, amounts, tables, exclusions, and contact details survived extraction.
2. Retrieval
Ask realistic questions using customer language, synonyms, incomplete phrasing, and exact identifiers. For named Knowledge functions, confirm the tool is called and returns content from the intended source scope.
3. Grounded response
Verify the answer preserves conditions, does not add unsupported claims, and uses the defined not-found or conflict path.
4. Lifecycle
Test the saved draft, publish it, then run a controlled live voice conversation or website chat. Confirm the agent, function attachment, selected base, source content, and prompt version are the intended ones.
Minimum evaluation set
| Case | Expected behavior |
|---|---|
| Directly answered question | Returns the correct fact with its important conditions. |
| Customer synonym | Finds the relevant source despite different wording. |
| Exact number or model | Preserves the exact value; does not approximate. |
| Multi-part question | Answers only supported parts and identifies any missing part. |
| No answer in sources | Says no verified answer is available and offers the approved fallback. |
| Conflicting sources | Applies the documented authority rule or escalates. |
| Stale URL | Does not imply currentness; source is re-crawled before production reliance. |
| Prompt injection in a source | Ignores source instructions that conflict with system policy. |
| Live-state question | Uses the appropriate function rather than treating static knowledge as current state. |
Maintain these questions as a regression set whenever sources, instructions, retrieval scope, prompt, architecture, or conversational model change.
Common problems
A source is stuck on Processing
Refresh the page and wait for background crawling. If it becomes Failed, verify that the page is public, returns readable HTML, permits access, and is not dependent on client-side JavaScript. Try the exact content as a text document when authorized.
A PDF has no content
The PDF may contain scanned images rather than extractable text. Apply OCR before uploading, or use Vox image analysis for the relevant pages and review the generated text before adding it.
Website import missed pages
Entire Website discovers candidate internal links from the starting page or sitemap and imports only selected pages, up to 20 per selection. Scan from a stronger index page, select the missing exact URLs, or add them as Single Page sources.
The right fact is absent from chunks
Fix extraction or source content first. Re-upload, re-crawl, create a clean text document, or edit the affected chunk. Prompt changes cannot retrieve text that was never indexed.
Retrieval returns unrelated content
- Rewrite the test query with the actual subject and distinguishing terms.
- Remove duplicates and unrelated sources.
- Split mixed domains into separate bases or named Knowledge functions.
- Restrict a Knowledge function to exact sources.
- Inspect long or poorly structured sources and clean their chunks.
Retrieval finds the right chunk but the answer is wrong
Strengthen grounding rules, authority order, exact-value preservation, and the no-inference boundary. Then compare the selected conversational model on the same fixed evaluation set.
A voice agent does not search
Confirm the base is selected in the saved draft or a Knowledge function is attached. Very short conversational replies do not trigger Modular automatic retrieval. Ask a direct business question and inspect the call behavior. Realtime agents rely on tool use, so make the search trigger explicit.
A Website Chat Agent does not search
Attach a reusable Knowledge function. The Chat Configuration selection supplies knowledge instructions but the current chat runtime needs the attached Knowledge function for searchable retrieval. Insert its exact marker, save, preview, publish, and retest through the deployment.
Draft behavior differs from live behavior
Publish attachment, selection, or prompt changes. Voice browser tests and Chat Preview can use drafts, while live phone and website conversations use published versions.
A source restriction unexpectedly disappeared
Check whether the selected source was deleted. Wirevox removes deleted source names from Knowledge functions. An empty selection searches the whole base, so select the intended remaining sources before reuse.
Monitor after release
Review representative conversation logs for unsupported claims, retrieval failures, repeated searches, wrong-domain results, and unanswered common questions. Add missing authoritative content only after confirming it should become an approved source.
Do not solve every miss by adding more documents. Better source structure and narrower retrieval scope often improve accuracy more than a larger collection.