Hi team. I am experiencing a persistent retrieval issue. My agent was previously reading and using an attached Knowledge Base perfectly. I deleted that initial KB and uploaded a new updated version. Since that deletion, the agent has completely stopped pulling any data from the new KB.
Here is what I have already tested to isolate the issue
Verified the conversational node is set to “Prompt” (LLM mode) and the new KB is properly linked to the node.
Formatted the text perfectly using Markdown with proper headings for chunking.
Created entirely new Knowledge Bases from scratch to rule out a single corrupted file.
Rewrote the global prompt to strictly force KB retrieval.
Despite all of this, the agent acts as if the KB is empty and defaults to its fallback prompt instructions. I suspect there is a backend vectorization error, a stuck status, or a caching bug preventing the agent from indexing the newly uploaded files.
Could you please check the backend logs for these IDs to see if the processing is stuck or throwing a hidden error?
Hi team.
I just tried again and the issue is still happening. To clarify the exact symptom, the agent is completely ignoring the attached Knowledge Base. Whenever I ask a question that is clearly answered in the KB, the agent immediately defaults to the fallback sentence I wrote in my global prompt which is “Bu detayı bölge yöneticimiz sizinle paylaşacaktır”.
This confirms the LLM is processing the prompt correctly, but the retrieval from the KB is failing or returning zero chunks.
The KB was working just fine, until i deleted it, and then created new KB. I created several KBs but none works.
Could you please look deeper into why the retrieval is failing for my KBs.
Hi Mark, is there a big problem we are facing now?
Is support working during weekends?
I’m so close to Demo presentation with my client. This is the last step. Looking forward to present on monday to my client. Will we fix by that time?
Thank you for looking into this, but the red box you highlighted is not proof that the Knowledge Base is functioning correctly. The phrase “Bir bin beş yüz” that appears right below your highlight is being pulled directly from the global prompt, not the Knowledge Base.
The core issue remains unresolved. The agent used to answer these exact same questions perfectly using the KB, and now it completely fails to retrieve that data.
This technical roadblock is entirely freezing my ability to publish this project. I need your team to help me get this agent working correctly today.
Since Retell is functioning fine globally but my specific agent is stuck in this retrieval error, let us take a proactive workaround. I want a clean start. I guess that will bring back the abilities I used to have in this agent.
Could your engineering team please take my existing flow and migrate it to a brand new database and a fresh agent instance.
Let us just put everything into a fresh backend environment so we can bypass whatever is corrupted in this current instance.
(In case of a bad stiuation, I have a copy of my global prompt and the KB text ready to go.)
I look forward to getting this reset and working today.
If you check the call’s history and click on ‘Knowledge base retrieval’, you can see that we retrieved KB chunks.
Could you check to confirm if either of the following are true:
Some relevant content of the KB are not retrieved.
The retrieved KB chunks are relevant, but they are not used by the agent.
