The answer boundary is written down

Knowledge-Grounded Chatbots

What makes a knowledge chatbot trustworthy is not how plausible an answer sounds. Every answer has to carry approved source support, audience access, and a defined path to clarification or silence.

An answer can sound plausible even when the source is unapproved or off-limits to the person asking. We build the chatbot around an approved source register, visible citations, access-aware retrieval, and clear rules for asking a follow-up or giving no answer.

The chatbot's answer flow cites its source, filters retrieval by the caller's identity, and gives no answer when no registered source supports the question.

Discuss the answer boundarySee the broader chatbot service
Illustration of Knowledge-Grounded Chatbots: a team shaping a conversational AI experience and its guardrails

Some of the 500+ brands we've worked with

See all references
  • Domino’s
  • Capital Dergisi
  • Halk Yatırım
  • Wall Street English
  • DLive
  • Tatilsepeti

Steps, gates, and who decides

How we work

We follow one question from caller identity to cited answer. The same review covers source approval, access, clarification, refusal, freshness, and conversation repair.

  1. Decide what can support an answer

    With source owners, we list the material the chatbot may use, its accountable owner, permitted audiences and channels, access rules, review needs, and the gaps that require clarification or silence.

    AI assist
    Working from the source set, the model groups candidate audience rules for owner review.
    Human gate
    Which sources enter the first answer boundary, and which audiences may use them. Each source owner approves the entry, audience, and freshness rule.

  2. Keep access attached

    We carry source and permission metadata through ingestion and indexing. Retrieval filters by caller identity, the answer cites its support, and an unclear or unsupported question follows the approved clarification or no-answer path.

    AI assist
    A model pass flags passages that lack a source or audience rule.
    Human gate
    Can each retrieved passage be traced to an approved source and an allowed audience? Identity filters and retrieval rules need the access owner's approval.

  3. Test answers and silence

    We run ordinary and unclear questions, cases with no support or stale material, and cross-audience attempts. Reviewers inspect claim support, citations, authorization leakage, no-answer choices, clarification, and conversation repair.

    AI assist
    Adversarial conversations drafted by the model probe unsupported answers and access leaks.
    Human gate
    Which failures block release and which remain visible as accepted limits. Every flagged measure failure goes to the source and access owners for acceptance or rejection.

  4. Hand over the source routine

    We tie source updates, review checks, no-answer handling, retention, approvals, and rollback to named people. The routine sets the next review and what happens when support, access, or freshness fails.

    AI assist
    The model drafts the first routine from open findings and source-update checks.
    Human gate
    Who owns freshness, the next review, release, and rollback. The source and service owners accept the duties, review cadence, and release limits.

Named artifacts you keep

What you get

The four records let reviewers trace source authority, audience access, answer behavior, and the work that follows a source change.

  • Matrix

    Source, audience, and freshness register

    For every source, it records the owner, approved audiences, access rule, intended use, and freshness requirement.

  • Architecture document

    Identity-filtered retrieval and citation map

    A working map from ingestion and indexing through identity filtering, retrieval, citations, conversation context, clarification, and abstention.

  • Dataset

    Support, access, and no-answer test report

    Representative questions with expected support, citations, access behavior, freshness, clarification, no-answer choices, and repair.

  • Playbook

    Freshness checks and reversible-release notes

    The agreed update checks, freshness owner, no-answer response, human verification, approval gates, retention, and reversible rollout.

Scope and honest limits

When to bring us in

This build works when source and access owners can draw the answer boundary, name its audience, and agree when the chatbot should stay silent.

A good fit when

  • Source owners can confirm which material is current, but the chatbot still lacks a register showing which audiences may rely on each source.
  • Audience rules exist for each channel, yet callers can retrieve passages without a trace back to the access decision that allowed them.
  • Real questions show unsupported answers, but no-answer and clarification choices still differ before the source owner releases the chatbot.
  • Approved sources have named owners, while freshness and audience rules still need to follow each item through ingestion and retrieval.
  • The chatbot returns citations, but identity filtering and the no-answer path can break when a question is unclear or unsupported.
  • Conversation tests cover ordinary questions, yet stale support and cross-audience leakage remain invisible until someone reads the trace.
  • A source owner can approve release, but verification gates and a reversible rollout are not recorded in one operating plan.

Better handled as other work when

  • You want the chatbot to answer from unregistered material. Every response instead needs support the caller is permitted to retrieve.
  • An automated score should replace source-owner judgment. Named access and release owners keep final approval.
  • You need Zeo to find and maintain source content or run the chatbot after release. That ongoing work requires a separate service scope.

If one of these is closer to your situation, start here instead: See the broader chatbot service

Engineers who ship production AI

This is the part of Zeo that writes and ships code. Our senior engineers build agents, chatbots, and RAG pipelines, along with the automation and data work around them, and they keep operating those systems once they're live. We've worked with more than 500 brands since 2011.

Content we've produced on this topic

Articleby Burak PehlivanWhat is WebMCP? Agentic Web and Websites Interacting with AI AgentsArticleby Can MutioğluOpenAI Models & APIs UsageArticleby Yiğit KonurMCP Server Architecture: State Management, Security & Tool Orchestration

Tools we use

Tools behind this work

  • LlamaIndex

    keeps every retrieved passage linked to its approved source

  • Weaviate

    combines semantic retrieval with access and source metadata filters

  • Sentence Transformers

    provides self-hosted embeddings for sensitive or specialized knowledge corpora

  • Ragas

    tests grounded answers, relevant retrieval, and correct silence behavior

  • Langfuse

    shows query, filters, passages, answer, citations, and evaluation together

  • Guardrails AI

    enforces citation, answerability, and response-shape rules before delivery

Next step

Bring the sources behind the answers

Provide the sources, access rules, and conversation examples. We'll define the smallest grounded slice worth testing.
Define the source boundary

Before you decide

Questions we hear a lot

First, the authoritative sources need named owners and audience rules. We also need representative questions, freshness expectations, no-answer cases, channel context, and the criteria used to judge the result. The people responsible for source authority, access, and release have to be available for those decisions.