AI Agent Doesn't Pull Up Information From Knowledge Base

Good day!

I’m currently training our AI agent and have uploaded our knowledge base. However, I’ve noticed a few issues:

• It doesn’t dig deeply enough into the content, so the answers come out vague.
• It doesn’t pull up the links included in the knowledge base.
• When I train it under “Teach Our Agent,” it seems to learn correctly, but when I test it through Gmail, it responds that the information is not in the knowledge base.

Could you advise if there’s anything we can do to improve its performance?

Hi Pam,

Thanks for reaching out to Jotform Support. Based on the information from the screenshot you provided, I think the issue in general was about how the AI Agent processed the information from the Knowledge Base. You can train your AI Agent to understand your business even better by adding specific details about your products, services, and processes. Adding information to your AI Agent's Knowledge Base is easy. Let me show you how to do it:

1. In Agent Builder, in the blue navigation bar at the top of the page, click on Train.

2. Now, click on Add New Knowledge.

Note: If you’ve already added knowledge to your agent, click on Add New Knowledge to open the knowledge-adding options.

3. Enter accurate details your AI Agent can use to provide the best responses. This includes your Company Overview, Product features, Pricing, Policies, FAQs, and any other key information customers may ask about.

4. Now, click on Save at the bottom of the page, and you're all set.

This process will ensure that the information will be added to your AI Agent's knowledge base properly. You might also want to check out our guides about How to Train Your AI Agent, How to Teach Your Agent by Chatting and How to Train Your Agent with Questions and Answers.

Reach out again if you need any other help.

Hi,
We have uploaded files with specific FAQs that your clients ask. We want the AI agent to use the exact wording that we give for each question. The agent is not following the guidelines we have given it. It is not accessing previous email responses even though we have connected it to our gmail account. I am not able to get helpful support from Jotform either. Is there a real support person that can screen share and show is what the problem is. Thank you,

Pam Bartha

Hi Pam,

Can you also share with us a screen recording of what’s happening so we can try to replicate it on our end? You can share the video recording on any public sharing platform like Google Drive or Dropbox, or you can simply use Loom to record the video and share the video link with us.

To thoroughly troubleshoot the issue you're experiencing, our team would like to clone your AI agent along with its associated knowledge base. By replicating your environment, we can more accurately diagnose the issue and work on a tailored solution. Rest assured that all data will be handled securely and used solely for troubleshooting purposes in line with our strict privacy and security policies. Once the issue is resolved, the cloned AI agent and all associated data will be deleted.

Can you confirm that we have your permission to go ahead with this process? If you have any questions or concerns, we’re here to help.

When we hear back from you, we’ll be ready to move forward.

I created a short video, but I don't know why I had to do that because you could just look at the thread of myself trying to train Evelyn and you will see that she is not answering the questions from our knowledge base that we gave her. She's making a lot of errors. Most of her responses are completely erroneous. I am really disappointed with the level of support. This has been gone going issue for days and I'm asked to make a video about it? There is no one that can screen share to help resolve the problem quickly? I honestly don't have the time to invest your AI tool that likely will not work well and the support is very poor. Unless you can help me quickly I am going to have to cancel my subscription. This has not been a good experience. I was really looking forward to working with your company to create various AI agents and forms, etc. but this is very frustrating and frankly a waste of my time.

Hi Pam,

I understand how important for you to fix this. In order to thoroughly troubleshoot the issue you're experiencing, our team would like to clone your AI agent along with its associated knowledge base. By replicating your environment, we can more accurately diagnose the problem and work on a tailored solution.

Rest assured that all data will be handled securely and used solely for troubleshooting purposes in line with our strict privacy and security policies. Once the issue is resolved, the cloned AI agent and all associated data will be deleted. 

Once you get back to us, we can take the next step toward solving this.

Hi, Yes you can clone my AI Agent. Thanks, Pam Bartha

Hi Pam,

The best approach to train the AI Agent to use the exact wording for each of your questions is to add them as Actions. By doing this, you can customize the trigger for when the AI Agent should use the specific wording you’ve set. This makes the responses more accurate, consistent, and easier for the AI to handle. It's really easy to do. Let me show you how:

1. In Agent Builder, in the blue navigation bar at the top of the page, click on Train. 

2. In the menu on the left side of the page, click on Actions. 

3. Set up the Action like you see below and click on Save:

Take a look at the screenshot below to see how it works after: Give it a try and let us know how it goes.

Hello!

I have tried your suggestion but it is still unsatisfactory.

Here's what I have tried
You can see that it still failed to answer word-for-word even though I have already given it the EXACT answer.

Also, if this is going to be the solution for it to be able to give verbatim answers, does that mean we have to input all possible questions and answers to the ACTION tab?

Your clarification would be very much appreciated.

Hi Pam,

You can use the Teach Your Agent feature to give a verbatim answer when a user asks something. Just set the exact response you want, and the agent will use that when the question comes up. Let me walk you through it:

1. In Agent Builder, in the blue navigation bar at the top of the page, click on Train.

2. In the menu on the left side, click on Teach Your Agent.

3. Then, just give the exact info you want your AI Agent to say when a specific question comes up.

4. Now, it'll automatically create the knowledge base according to the information you provided. That's it.

Here's the result:

Note that this feature sticks to whatever instructions you give it. If you want the AI Agent to give answers exactly as the user asks, just add those details when setting it up, and it'll follow them word for word.

Give it a try and let us know if you need any help.

Hello !

We have been training our AI Agent and would like our questions answered.

What kind of things we could do to help the AI agent be more accurate with the knowledge base we want it to have? Are there any special parameters for setting up her knowledge base so that the AI agent will use it more accurately?

Your assistance would be greatly appreciated.

Hi Pam,

We appreciate all the updates and continued cooperation from your end. I understand that getting your AI agent to respond exactly as intended is key. Here are some additional strategies that you could try which should improve your agent's efficiency and accuracy:

Use clean Q&A entries in the Knowledge Base

• Format each entry as a short, precise question → answer pairing, rather than long paragraphs. This helps the AI agent match user inputs more reliably.
• If a user’s phrasing deviates slightly, synonyms or alternate phrasings (added in Q&A) help the match.

Leverage Actions for your must-say responses

• For critical questions (compliance, legal, or branding), store those exact responses in an Action rather than relying on the generative AI logic.
• In Action triggers, use the same keywords or phrases that users tend to say so the agent picks the correct scripted response.

Add alternate wording (variants/synonyms) for each question

• For example, if the canonical question is “What is the refund policy?” also add variants like “How do refunds work?”, “Return rules”, etc.
• This helps the agent catch that intent and pick the correct answer more reliably.

Short, modular content wins

• Keep entries concise and avoid nested, complex structures. One fact per item, bulleted format, or short sentences helps the AI pick the right piece.
• Users in Jotform forums confirm that overly complex or nested information often “confuses the agent” and reduces its accuracy.

Regular testing & iteration

• Use real user questions (from support tickets, email logs) to test whether the agent matches correctly.
• If a test fails, tweak the question phrasing in the KB or Action and retry.
• Periodically review and refine your KB entries to keep them aligned with real usage.

Adding these best practices should help improve the performance of your AI agent dramatically.

Hoping this was helpful. If you have further questions or concerns, don't hesitate to reach back out.

I just want to know if we do that, would it be able to get answers from our KB word-for-word? We have been training it to do that but until now, even after countless alterations and trainings, it is still not able to do that.

Hi Pam,

To make sure that the AI Agent will answer using the imported Knowledge Base, you can still work with the Action feature under the Train tab. Instead of typing the words you'd like the AI Agent to answer when a specific question is asked, you can choose the imported Knowledge Base to show. Let's walk through this together using a cloned one:

1. In AI Agent Builder, on the blue navigation bar on top, go to the Train tab.
2. Then select Actions from the left menu, and click on the Add new action button on top.
3. On the DO section, click on the downward arrow icon, scroll down and choose Answer Using Knowledge Base.
4. Then you can select which Knowledge Base you'd like to show. Finally, click on the Save button.
   

Keep us updated and let us know if you require additional help.

1. LDF DIET FAQ 

-This file is the sole source of truth. The agent must provide answers strictly word-for-word from the file and must not rephrase, summarize, or invent content. If the file does not contain the requested information, the agent should indicate that the answer is not available.

• This is the exact prompt that I gave Evelyn when I uploaded our Diet FAQs. Please help me understand why she still cannot follow this. Note that there are no other files uploaded yet as we are testing her ability to answer FAQs about out eating plan. She would either come up with her own answers, add/remove info from the answers in the FAQ file, or say that it is not included in her knowledge base when the questions I am asking her are all from the FAQ.

Hi Pam,

Thanks for providing additional details.

I understand you're intending for the AI agent to abide by a strict word-for-word answer from a specific FAQ file uploaded to it. This should definitely be doable, but I believe you may have to start trying to force the agent to use scripted Actions instead of just the knowledge base references.

Here is what I would suggest trying:

I would recommend trying to create a brand new agent for testing purposes that way you can compare how each one behaves according to your testing methods.

1) Make Actions the source of truth (not the KB)

Go to TRAIN → Actions → + New Action.

1. WHEN: choose one of:

• User asks about → paste the exact question, plus a few common variants (one per line).
• or The sentence contains → add key trigger phrases for that question.

1. DO: Say exact message → paste the verbatim answer from your Diet FAQ.
2. Save. Repeat for each FAQ that must be word-for-word.

Why this works: Actions return exactly what you write and run before the general knowledge behavior. The regular Knowledge Base is still useful for non-critical topics, but it’s intentionally generative and can rephrase. Actions won’t.

2) Tell the agent to not paraphrase

TRAIN → AI Persona (Guidelines):

• Add rules like:
• “For any topic covered by the Diet FAQ, respond only with the exact answer from the corresponding Action. Do not rephrase, summarize, or combine answers.”
• “If no Action matches, say: ‘This answer is not available in my Diet FAQ.’”

1. TRAIN → Actions: add a final fallback Action:

• WHEN: User asks about (general catch-all, e.g., “diet plan”, “diet faq”, etc.)
• DO: Say exact message → “This answer is not available in my Diet FAQ.”

3) Reduce or disable generative fallbacks that cause drift

In TRAIN → Actions, ensure your Q&A Actions are listed above any “Answer using knowledge base” Actions so they take priority.

• If you added an Action like Answer using knowledge base, restrict its WHEN so it doesn’t overlap with your FAQ triggers.
• If you used Teach Your Agent or a text Knowledge Base for the same FAQ, keep it for reference but rely on Actions for critical verbatim answers.

4) Make the FAQ machine-readable (if you still want KB as a backup)

If you keep the Diet FAQ in the Knowledge Base:

• Use text-based entries or a clean .docx/.txt (avoid scanned PDFs or image-only PDFs).
• Structure as Question on one line + Answer in a short block (bullets are fine).
• Split long FAQs into multiple short Q&As rather than one giant document.
• Use the exact question wording you expect users to ask, and add 2–3 common variants.

5) Channel consistency (your Gmail test vs. builder)

Sometimes people test in Gmail before republishing:

• After updating Actions/Persona, Publish the agent again.
• In PUBLISH → Gmail, make sure the same agent is attached to the Gmail channel. If you have multiple versions, remove/re-add to ensure Gmail uses the latest.
• Test first in Standalone preview, then in Gmail.

I understand that this is a lot of information and steps to consider, but I believe it would be a great start to learning more about how your first AI Agent is behaving compared to another one based on certain settings and conditions.

We have been training our AI agent in Jotform using ‘Actions’ as the source of truth, where we manually enter each FAQ question along with its corresponding answer. This setup works well for providing verbatim responses from the FAQ. However, I’m concerned about how the AI will handle questions related to the LDF diet that are either not included in the FAQ or are phrased differently by users. How can we ensure the agent can respond accurately in those cases?

Hi Pam,

You can keep using Actions as the main source for FAQs where exact wording matters. This ensures the AI provides accurate and consistent responses for those specific questions. For anything outside the FAQ or asked in a different way, the AI can rely on your wider Knowledge Base, which allows for more flexible and conversational responses.

To improve matching accuracy, make sure to include common variations or synonyms of each question in both your Actions and Knowledge Base entries. It also helps to keep an eye on real user questions so you can update your content and fill in any gaps or add new phrasing you might not have covered yet. Lastly, you can clone an existing AI Agent in your account and test it further with the cloned form to see if you can replicate the issue. Let me show you how:

1. On your My Workspace page, hover your mouse over your form and click on More, on the right.
2. In the dropdown menu, click on Clone.
3. In the window that opens up, enter the Agent Name and click on Clone Agent.

Let us know if there’s anything else we can help you with.