Skip to content
Mainframe
LOCAL SEARCH

Search Mainframe

    Start a project

    Mainframe Lab 02 · Reference architecture

    Internal Knowledge Assistant

    Make documents searchable, support answers with sources, and keep uncertainty visible—without claiming hallucinations are impossible.

    RAGVector searchPermissionsEvaluation

    System map

    Reference architecture flow

    Reference architecture flow
    1. 01ApprovalAllowed documents and permissions
    2. 02PreparationVersions, chunks, metadata
    3. 03RetrievalSemantic and filtered matches
    4. 04AnswerContext, sources, uncertainty
    5. 05ReviewTests, feedback, handover

    Starting point

    The problem this build examines

    Knowledge is spread across manuals, policies, and project folders. Search takes time, while a general chatbot may produce answers that sound plausible but lack evidence.

    Approach

    How the system is structured

    Approved documents are versioned and split into citable sections. Retrieval supplies relevant excerpts; the model drafts an answer and displays its sources. With weak evidence, it declines or hands the question to a person.

    Evidence

    What this concept build demonstrates

    How sources, permissions, representative test questions, and a controlled “no answer” route work together.

    Acceptance checks

    • Unapproved documents remain excluded
    • Every answer shows the sources it used
    • Representative questions are tested against expected evidence
    • Weak retrieval leads to clarification or handoff

    Typical handover

    • Document and permission model
    • Ingestion and retrieval pipeline
    • Evaluation set with representative questions
    • Operating, update, and deletion process