Knowledge Base

A knowledge base is a set of documents you upload to your organization that a voice agent can search while a call is in progress. Upload a PDF, Word document, or text file once, and any conversation step you attach it to can pull relevant passages from it to answer a caller's question.

Behind the scenes, Lemu splits each document into chunks, generates an embedding for every chunk, and stores them so they can be searched by meaning rather than exact keywords — the agent can find the right passage even if the caller phrases their question differently from the document's wording.

Uploading documents#

Documents live under Files in the dashboard (https://app.lemuai.com/files), reachable from the Configuration section of the sidebar.

The Upload tab accepts:

• File types: PDF, DOCX, DOC, TXT
• Max size: 5 MB per file

Drag a file onto the upload area or use Choose File. Once the file finishes uploading, processing starts automatically — the All tab lists every document with its filename, size, chunk count, and processing status, polling for updates every few seconds while a document is still processing.

Only documents with status completed can be attached to a conversation step — a document that's still processing or failed won't show up as selectable.

You can delete a document from the All tab at any time; this also removes its chunks from search results.

Using a knowledge base in a voice agent#

Knowledge base documents are attached per conversation step, not globally to the whole agent. Open the Start Call node or an Agent Node in the workflow builder, and scroll to the Documents section below Tools — it lists every completed document as a checkbox so you can select which ones this step can reference.

Agent Node
├── Prompt
├── Tools           (HTTP API tools, etc.)
└── Documents        ← knowledge base documents for this step
    ☑ refund-policy.pdf
    ☐ product-catalog.docx

If a step has no documents selected, the agent has no knowledge base access on that step. If it has one or more, Lemu registers a retrieve_from_knowledge_base function with the LLM for the duration of that step, scoped only to the documents you checked.

What happens during a call#

The LLM itself decides when to call the tool — based on your node prompt and the conversation so far — and what search query to send. It isn't a fixed keyword match: write your prompt so the model understands when a caller's question should be looked up (for example, "if the caller asks about pricing, policies, or anything that might be in the uploaded documents, use the knowledge base tool before answering").

A few things worth knowing about the search itself:

• Each call returns a small number of the most relevant chunks (three, by default) — not the whole document.
• Results aren't filtered by a similarity cutoff before reaching the LLM; the model sees the best matches available and judges relevance itself.
• The search is always scoped to your organization and to the documents selected on that node — a document isn't visible to a step unless it's checked on that step.

Searching documents directly#

The same vector search that powers the in-call tool is also available as an API endpoint, so you can query your knowledge base outside of a live call — see the Knowledge Base section of the API reference for request and response details.