Skip to content
Zumu
Demo

Knowledge base for AI agents

Your agent is only as good as what it knows three seconds in.

Point Zumu at your documents, call recordings, and website. It builds a knowledge graph, not a pile of embeddings, and validates every claim through adversarial multi-model debate before your agents query it live, mid-call.

  • Adversarial validation before an answer ships
  • A graph, not a pile of embeddings
  • Audio and video become knowledge
  • Answers cite their sources
Book a demoWatch it build the graph
Pipeline stages, purpose to graph
8
Ingestion depths, light to research
4
File types, plus a crawl of your own website
11 +site

Pipeline structure, not a performance claim. Every stage is described below.

A source column, eight pipeline stages, and a graph that grows while you watch.

An operations manager sorting through a stack of paper policy documents at her desk

Somewhere there is a binder nobody can query.

How it works

Four jobs, one graph

Each one traces to a real stage in the pipeline. None of them are a marketing gloss over a vector database.

  • Ingestion

    Point it at everything you already have

    Drop files, point it at your website, or upload a folder of call recordings. Four ingestion depths trade speed for thoroughness, from a light pass to a research mode that reaches for a stronger model on extraction.

    • .txt, .md, .pdf, .docx, .html, .csv, .mp3, .wav, .m4a, .mp4, .webm, plus SRT/VTT subtitles
    • Call and video audio transcribed with Whisper before extraction starts
    • Website crawl at onboarding, capped so it finishes before the first session ends
  • The debateThe hero stage

    Nothing ships without an argument first

    One model drafts a claim from the source material. A second model reviews it, and a claim that conflicts with something already known gets struck through and rewritten before it reaches the graph.

    • Six-factor confidence scoring gates what survives into the graph
    • A privacy shield anonymizes PII before any model sees the source text
    • Checkpoint and resume, with per-session cost tracked against the run
  • Retrieval, mid-call

    The right store answers, not whichever one is loudest

    Retrieval is hybrid: dense vector search and BM25 keyword search, fused by reciprocal rank. A graph-aware path is tried first, so a question about how two things relate does not have to get lucky on a chunk.

    • Two-tier lookup: synthesized graph nodes first, raw source chunks second
    • 1024-dimension embeddings, fused rather than chosen between
    • The agent cites the source it pulled from, in the answer it gives the caller
  • The loop

    The graph gets better without you touching it

    A call scored as exemplary gets pushed back into the knowledge base as training material automatically. Nodes can be enriched, merged, reassessed, or promoted by hand when a person wants the final say.

    • Push modes: automatic, manual review, or batched
    • Every knowledge base carries its own quality score
    • A pipeline dashboard shows processing health across every ingestion job

Onboarding

Give us your URL. Your first agent knows your business before you finish signing up.

Onboarding crawls your own site while you are still filling out the rest of the form and builds your first knowledge base from it. There is a real knowledge graph waiting by the time you get to your first test call, not an empty one you have to fill yourself.

The crawl is capped so it finishes fast. Everything it finds still goes through the same debate before it counts as known.

Start free trial

How the graph stays consistent

3

Stores kept in sync on every write: Postgres canonical, Neo4j for relationships, Qdrant for the vectors. A failed write rolls all three back together.

6
Confidence factors scored on every claim before it can reach the graph
1024
Dimensions per embedding, fused with keyword search rather than chosen instead of it

Questions

Knowledge and retrieval, answered plainly

It is the system your agents query mid-call for anything beyond the conversation in front of them: policies, product details, past calls, whatever you have written down or said out loud. Zumu builds it as a knowledge graph rather than a pile of embeddings, so the agent can follow a relationship between two things, not just find a chunk of text that sounds similar to the question.

Book a demo

Give your agents something worth citing

Bring a document you would not trust a search bar with. We will show you the debate that has to happen before it becomes something your agent can say to a caller.

Book a demoTalk to sales
  • No credit card for the demo
  • Bring a document, watch it become a graph
Zumu | Knowledge Base for AI Agents, Built as a Graph