Knowledge Base
A knowledge base your voice agent searches mid-call
Upload your policies, price lists and FAQs as PDF, Word, text or Markdown. Attach them to a step of the call, and the agent looks up the passage it needs when a caller asks. Nothing is trained.

What is a knowledge base for an AI voice agent?
A knowledge base is a set of your own documents that a voice agent searches during a call, so it answers from your facts instead of from the model's memory.
This is called retrieval-augmented generation, or RAG: the model retrieves a passage first, then answers from it. In Bananaflow it is a set of files you attach to each step of the agent.
The idea was named in a 2020 paper, Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In plain words: retrieval augmented generation, and how it grounds an answer in your documents.
New to voice agents? Read how a voice agent call works, one turn at a time.
How does the Knowledge Base work?
You upload a file, pick whole file or search, attach it to a step of your agent, and the agent looks it up whenever a caller asks something the file covers. Six steps.
Upload a file
Open the Files page and drop in a PDF, Word, text or Markdown file. Each file can be up to 5 MB.
Pick a mode for the file
Choose Full Document for short files, or Chunked Search for long ones.
Wait for Completed
Each file shows pending, processing, completed or failed, and a failed file tells you why.
Attach it to a step
Add the file to the Start or Agent step that needs it. That step can now look it up.
Tell the agent when to look things up
Add one line to that step's prompt, such as "look up charges before you answer, and do not guess".
Test it, then publish
Try it in the browser, check the lookups in the transcript, and publish the agent.
Have a pile of policy documents to sort out?
What can the agent look up?
Anything written down in a PDF, Word, text or Markdown file up to 5 MB, from a one-page price list to a long policy. Four things make that work on a live call.
Read the whole file, or just the right passage
Pick the mode for each file when you upload it.
Rate card.md: whole file. Loan policy.pdf: search.
- Full Document hands the agent the whole text, which suits menus, price lists and short FAQs.
- Chunked Search splits a long file into short passages, each with the headings it sits under, and returns the 3 closest.
- Scanned PDF pages and tables are read too.
Each step reads only its own files
Attach files to a Start or Agent step, not to the whole agent.
Step "Loan questions": Foreclosure policy.pdf
- A billing step never pulls in the sales brochure.
- The agent searches only the files attached to the step it is on.
- Build with AI can attach files for you while it draws the flow.
The agent looks it up when it needs to, and you can see it
A lookup is a tool the agent calls, not text pushed into every turn.
"foreclosure charge after 12 EMIs" returns 3 passages
- Each lookup shows in the call transcript with the query the agent wrote and the passages it got back.
- Open a call and check the answer against the passage in seconds.
- Ask something the files do not cover, and the passages show what the agent had to work with.
Change a text file, and calls switch over when it is ready
Edit text and Markdown files in the browser and save.
FAQ.md saved. Processing. Completed.
- Live agents keep using the old version until the new one is processed, with no republish.
- Calls started in the meantime still get an answer.
- A PDF or Word file is replaced by uploading the new one and attaching it in place of the old one.
Files can be up to 5 MB each. Website import is not available today: save the page as a file and upload it.
How does a call center use the Knowledge Base?
A loan support desk stops hard-coding charges into prompts and keeps them in one file the agent reads. This example is made up: the names and numbers are illustrative, not a customer.

Monday 10:00 am
Two files go in
A two-wheeler loan desk in Jaipur takes about 1,800 calls a day. Ops lead Neha Sharma uploads Charges FAQ.md as Full Document and the 38-page Foreclosure policy.pdf as Chunked Search. Developer Rohit Jain attaches the FAQ to the Start step and the PDF to the Loan questions step, adds "look up charges, never guess" to its prompt, tests by text chat and publishes.
Monday 2:15 pm
A caller asks about closing the loan
A caller asks what it costs to close the loan after 12 EMIs. The agent looks it up, gets the passage under "Foreclosure after 12 months" and answers: 4% of the principal still owed. Neha opens the call and sees the query and the passage in the transcript.
Thursday
The charge changes
The lender cuts the charge to 3% from 1 November. Neha edits one line in Charges FAQ.md and saves. Calls keep the old answer for the minute it takes to process the file, then use the new one. No prompt change, no republish.
Support teams use the same idea to keep answers consistent. See how support teams stop wrong answers.
How does the Knowledge Base work with the rest of the agent?
It is one tool among the agent's others. Three places it meets the rest of your setup.
Files sit on the steps of your workflow
In the workflow builder, add files to the Start or Agent step that needs them. Each step reads only what you attached to it.
Files for what is written down, tools for what changes
Put what is written down and the same for every caller in a file. Use an HTTP tool for anything that changes per caller, such as an order status.
Lookups show in the call transcript
Every lookup is a row in the call's transcript, with its query and passages. The same transcript is what Call QA reads after the call ends.
Set up files per step in the workflow builder, and see how each call is reviewed after it ends.
Workflow builder
Attach files to the step of the call that needs them.
Call QA
How each call is reviewed after it ends.
AI customer support calls
Troubleshooting steps the agent gives on every call.
AI inbound calls
Inbound calls answered in your callers' languages.
Telephony and SIP
The phone lines your agent answers and calls on.
How a voice agent call works
One turn at a time, from the caller's last word to the reply.
FAQ
Knowledge Base questions, answered
What is a knowledge base for an AI voice agent?
A knowledge base is a set of your own documents that a voice agent searches during a call, so it answers from your facts instead of from the model's memory. In Bananaflow it is a set of files you attach to each step of the agent.
Do you train the AI on our documents?
No: nothing is trained, your files are split into passages and searched when the agent needs them. We do not use your call data to train or fine-tune AI models.
Which files can we upload?
PDF, Word (DOCX), plain text and Markdown files, up to 5 MB each. Scanned PDF pages and tables are read too.
Can the agent read our website?
Not directly today: save the pages you need as a PDF or text file and upload them.
How do we update a document?
Edit a text or Markdown file in the browser and save it, and calls use the new text once it has been processed. For a PDF or Word file, upload the new version and attach it in place of the old one.
How do we stop the agent from guessing?
Tell it in the step's prompt to answer only from what it looks up and to offer a transfer when the file does not cover the question, then check the lookups in the transcripts.
Should a fact go in the knowledge base or in a tool?
Put what is written down and the same for every caller in the knowledge base, and anything that changes per caller, such as an order status, in an HTTP tool.
Bring us one process. We'll pilot it on your real calls, with your team as the fallback.
Plan a pilot