Agent not responding correctly
Diagnosis for when the agent gives wrong, generic, or unfound responses even though the information is in your content.
If the agent is responding but the answers are incorrect, generic, or "I couldn't find that information" when the information exists in your content, follow this diagnosis.
Understand how the agent responds
Before diagnosing, it's important to know that the agent does not memorize your content — it searches the most relevant passages in real time (RAG) and generates the response based on them.
When the agent says "I couldn't find it," there are two possible causes:
- The information isn't in the indexed sources
- The information is there, but the search didn't retrieve the correct passage
1. Verify the information is indexed
Dashboard → agent → Training → click the suspected source → View chunks.
Search for a keyword from the expected information. If the chunk appears, the content is indexed. If not, the source wasn't indexed correctly.
Solution: Click Re-index on that source, wait for it to complete, and test again.
2. Check chunk quality
Even if the content is indexed, the chunk may be poorly split. Examples:
- Tables cut in the middle
- Answers spread across different pages with no semantic connection
- Headings without their corresponding content
What you can do:
- For PDFs with tables: rewrite the table as a list in the Direct text field — the agent understands lists better than complex tables
- For sites with content spread across multiple pages: add each URL individually instead of doing a full domain crawl
3. The question is very different from the content
RAG searches by semantic similarity. If the user asks "do you accept Pix?" but your content says "instant bank transfer available," the search may not find it.
Solution: Add a text source with synonyms and alternative phrasings:
Accepted payment methods:
- Credit card
- Debit card
- Bank transfer (ACH / wire)
- PayPal
4. The agent is "making things up"
If the agent is responding with things that aren't in your sources, check:
Behavior settings: Dashboard → agent → Behavior → enable "Respond only based on sources".
With this option enabled, the agent responds "I don't have that information" for out-of-scope questions instead of trying to answer using general knowledge.
5. Slow responses (more than 5 seconds)
High latency can indicate:
| Cause | Diagnosis |
|---|---|
| Too many sources to search | More than 500 chunks in the agent → consider splitting into 2 agents |
| LLM under load | Check status.simple-agent.me |
| User's connection | Test on a different network |
6. Responses in the wrong language
The agent automatically detects the question's language and responds in the same language, as long as the content is available in that language.
If the user asks in English but all content is in another language, the agent may mix languages.
Solution:
- Add sources in the desired language, or
- Under Behavior → Fixed language, force a specific language
7. The agent isn't using recent information
If you updated your site content but the agent still uses the old version:
- Dashboard → Training → select the source
- Click Re-index now
- Wait for the status to change to
Active
The embedding cache is cleared with each re-index.
Using the Playground for diagnosis
The Playground (tab on the right in Agent Builder) is the best diagnostic tool. Use it to:
- Replicate the exact question the user asked
- Click View sources used on the response — this shows which chunks were used
- If the chunks are correct but the response is wrong, the problem is in generation (adjust the system prompt)
- If the chunks are wrong or missing, the problem is in indexing or question phrasing
Widget not loading → · Webhook failing → · Re-index sources →