Treewalker starts two concurrent sessions—one probing ungrounded parametric memory with token-level logprobs and another using live Google Search grounding—then aggregates both over multiple runs to generate a holistic cross-comparison analysis of ranking discrepancies and blind spots.

sequenceDiagram
    autonumber
    actor User
    participant TW as Treewalker Engine
    participant UG as Ungrounded (Logprobs)
    participant G as Grounded (Live Search)
    participant DB as Postgres DB
    participant H as Holistic Analyzer

    User->>TW: Start paired session
    TW->>DB: Initialize session pair
    par Memory vs Live Search
        TW->>UG: Tree-walk alternative branches
        UG-->>TW: Token logprobs & paths
    and
        TW->>G: Run grounded queries
        G-->>TW: Live search associations
    end
    TW->>DB: Save branch items & run data
    TW->>H: Compare ungrounded vs grounded
    H-->>DB: Store gap insights & metrics
    DB-->>User: Return paired report