If the Dragwyb Click to Chat AI assistant gives inaccurate, incomplete, or off-topic answers, check whether the right knowledge-base content is being retrieved before changing prompts or models. The assistant’s response depends on its configured instructions, selected provider/model, and the source context supplied for the question.
1. Test with a known answer
- Choose a question whose answer is clearly stated in a published page or other source included in your knowledge base.
- Use a short question with the same key terms used by the source.
- Compare the assistant’s response with the source content, not just with what you expected it to know.
- Repeat with another known question to see whether the problem is isolated to one topic or affects all responses.
2. Check which source content is retrieved
The plugin’s retrieval flow uses indexed documents and embeddings to select context for the AI response. A page can exist on the website without being present in the assistant’s indexed context, and retrieval may return content that is only loosely related to the question.
- Open the Knowledge Base area and confirm the intended source is included in the indexing scope.
- Check whether indexing and embedding generation have completed and whether the configured embedding/vector-storage providers are working.
- Review any source links or retrieval details exposed by the assistant to see whether the relevant page was selected.
- If the wrong source is retrieved, improve the source’s headings and terminology, check the indexing configuration, and re-index where the plugin provides that action.
3. Review assistant instructions and answer boundaries
Open the AI Assistant settings and review the configured system instructions or prompt. Make sure they describe the assistant’s role and the desired answer style without contradicting the information in the knowledge base. The inspected source instructs the model to answer from the provided knowledge-base facts and acknowledge missing or contradictory information rather than inventing details.
- Keep instructions specific and consistent with the support or store use case.
- Do not ask the assistant to answer only from the knowledge base while expecting it to supply facts that are absent from the indexed content.
- Review custom prompts for conflicting rules, outdated product details, or overly broad claims.
- When information is missing, prefer a clear limitation or supported handoff to inventing an answer.
4. Verify the selected provider and model
Confirm which response provider and model are configured, that the API key is valid, and that the account has access to the model. If poor answers began after changing providers or models, test the same known question with the prior working configuration if it is still available and appropriate.
5. Check errors and request behavior
- Review AI error logs when logging is enabled, and redact secrets or customer data before sharing them.
- Check browser/network errors if the response is cut off, missing, or not displayed even though a request was sent.
- Test with a short prompt and without sensitive customer information to separate retrieval problems from provider or UI failures.
- Do not publish API keys, full private conversations, guest ticket tokens, or unredacted logs when asking for help.
6. Apply a controlled fix
- Correct the knowledge-base source or indexing scope first if the wrong context is retrieved.
- Adjust the assistant prompt only when the instructions are the cause of the mismatch.
- Change the provider or model only after recording the current settings and testing a comparable question.
- Repeat the same test questions and compare responses against the verified source content.