Skip to Content

Monolithic Canonical Pipeline Internal (v2.6.33)

Last modified by Robert Schaub on 2026/02/08 08:31
Information

Current Implementation (v2.6.33) - Single-context LLM analysis with tool calls, producing canonical schema output. File: monolithic-canonical.ts (1100 lines)

Monolithic Canonical Pipeline Internal Flow


flowchart TB
    subgraph Entry[Entry Point]
        INPUT[runMonolithicCanonical]
        BUDGET[Initialize Budget maxIterations 5 maxSearches 8]
        TIMEOUT[Start timeout 3 minutes]
    end

    subgraph Phase1[Phase 1: Claim Extraction]
        EXTRACT[LLM Call ClaimExtractionSchema]
        SCOPES[detectScopes heuristic fallback]
        CLAIMS[Build subClaims array]
    end

    subgraph Phase2[Phase 2: Research Loop]
        SEARCH[searchWebWithProvider]
        FETCH[fetchSourceContent parallel]
        FACTS[LLM Call FactExtractionSchema]
        PROVENANCE[filterFactsByProvenance]
        ITER_CHECK{Iteration limit?}
    end

    subgraph Phase3[Phase 3: Verdict Generation]
        VERDICT[LLM Call VerdictSchema]
        WEIGHT[calculateWeightedVerdictAverage]
        HARM[detectHarmPotential]
        CONTEST[detectClaimContestation]
    end

    subgraph Transform[Schema Transform]
        CANONICAL[Transform to canonical schema]
        REPORT[Generate markdown report]
    end

    subgraph Output[Output]
        RESULT[resultJson canonical]
        MARKDOWN[reportMarkdown]
    end

    subgraph Fallback[Fallback on Failure]
        FALLBACK[Fall back to Orchestrated]
    end

    INPUT --> BUDGET
    BUDGET --> TIMEOUT
    TIMEOUT --> EXTRACT

    EXTRACT --> SCOPES
    SCOPES --> CLAIMS

    CLAIMS --> SEARCH
    SEARCH --> FETCH
    FETCH --> FACTS
    FACTS --> PROVENANCE
    PROVENANCE --> ITER_CHECK
    ITER_CHECK -->|No| SEARCH
    ITER_CHECK -->|Yes| VERDICT

    VERDICT --> WEIGHT
    WEIGHT --> HARM
    HARM --> CONTEST
    CONTEST --> CANONICAL

    CANONICAL --> REPORT
    REPORT --> RESULT
    REPORT --> MARKDOWN

    EXTRACT -.->|Schema validation fail| FALLBACK
    VERDICT -.->|Schema validation fail| FALLBACK

Key Characteristics

 Feature  Description
 Single Context  All analysis in one LLM conversation context
 Structured Output  Uses AI SDK Output.object with Zod schemas
 Canonical Schema  Output compatible with existing Jobs UI
 Fail Closed  Falls back to Orchestrated pipeline on validation failure
 Budget Enforcement  Hard limits on iterations, searches, fetches, time

Budget Configuration

 Parameter  Value  Purpose
 maxIterations  5  Maximum research iterations
 maxSearches  8  Maximum web searches
 maxFetches  10  Maximum source fetches
 timeoutMs  180000  3 minute hard timeout

Zod Schemas Used

 Schema  Purpose
 ClaimExtractionSchema  Extract mainClaim, claimType, searchQueries, subClaims, detectedScopes
 FactExtractionSchema  Extract facts with source attribution
 VerdictSchema  Generate verdicts with truthPercentage, confidence

Shared Modules Used

  • scopes.ts - detectScopes, formatDetectedScopesHint (heuristic fallback)
  • aggregation.ts - calculateWeightedVerdictAverage, detectHarmPotential, detectClaimContestation
  • claim-decomposition.ts - normalizeClaimText, deriveCandidateClaimTexts
  • provenance-validation.ts - filterFactsByProvenance
  • budgets.ts - createBudgetTracker, getBudgetConfig, recordLLMCall

Differences from Orchestrated

 Aspect  Orchestrated  Monolithic Canonical
 LLM Context  Multiple calls, fresh context  Single conversation context
 Orchestration  TypeScript code controls flow  LLM decides next steps
 Research  Iterative decide-search-extract  Tool loop with budget limits
 Schema  Built-in TypeScript types  Zod schema validation
 Fallback  None (default)  Falls back to Orchestrated on failure
 Code Size  9000 lines  1100 lines