Skip to Content

Orchestrated Pipeline Internal (v2.6.33)

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

Current Implementation (v2.6.33) - The default pipeline with TypeScript-orchestrated staged workflow. File: orchestrated.ts (9000 lines)

Orchestrated Pipeline Internal Flow


flowchart TB
    subgraph Entry[Entry Point]
        INPUT[runFactHarborAnalysis]
        NORM[Input Normalization]
        BUDGET[Initialize Budget Tracker]
    end

    subgraph Step1[Step 1: Understand]
        UC[understandClaim]
        LLM1[LLM Call]
        SCOPE[detectScopes heuristic]
        CLAIMS[Extract SubClaims]
        QUERIES[Generate researchQueries]
    end

    subgraph Step2[Step 2: Research Loop]
        DECIDE[decideNextResearch]
        SEARCH[searchWebWithProvider]
        FETCH[fetchSourceContent parallel]
        EXTRACT[extractFacts per source]
        LLM2[LLM Calls]
        BUDGET_CHECK{Budget exceeded?}
    end

    subgraph Step3[Step 3: Verdict Generation]
        CLAIM_V[generateClaimVerdicts]
        ARTICLE_V[generateArticleVerdict]
        LLM3[LLM Call]
        WEIGHT[calculateWeightedVerdictAverage]
        CONTEST[detectClaimContestation]
        HARM[detectHarmPotential]
    end

    subgraph Step4[Step 4: Summary]
        SUMMARY[generateTwoPanelSummary]
    end

    subgraph Step5[Step 5: Report]
        REPORT[generateReport markdown]
    end

    subgraph Output[Output]
        RESULT[AnalysisResult JSON]
        MARKDOWN[Report Markdown]
    end

    INPUT --> NORM
    NORM --> BUDGET
    BUDGET --> UC

    UC --> LLM1
    LLM1 --> SCOPE
    SCOPE --> CLAIMS
    CLAIMS --> QUERIES

    QUERIES --> DECIDE
    DECIDE --> SEARCH
    SEARCH --> FETCH
    FETCH --> EXTRACT
    EXTRACT --> LLM2
    LLM2 --> BUDGET_CHECK
    BUDGET_CHECK -->|No| DECIDE
    BUDGET_CHECK -->|Yes| CLAIM_V

    CLAIM_V --> WEIGHT
    WEIGHT --> CONTEST
    CONTEST --> HARM
    HARM --> ARTICLE_V
    ARTICLE_V --> LLM3

    LLM3 --> SUMMARY
    SUMMARY --> REPORT
    REPORT --> RESULT
    REPORT --> MARKDOWN

Key Functions

 Phase  Function  Purpose  LLM Calls
 Entry  runFactHarborAnalysis  Main entry point  0
 Entry  normalizeYesNoQuestionToStatement  Input neutrality normalization  0
 Step 1  understandClaim  Claim extraction and classification  1
 Step 1  detectScopes  Heuristic scope pre-detection  0
 Step 2  decideNextResearch  Generate search queries  1 per iteration
 Step 2  searchWebWithProvider  Web search (Google CSE or SerpAPI)  0
 Step 2  fetchSourceContent  Parallel source fetching  0
 Step 2  extractFacts  Extract facts from sources  1 per source
 Step 3  generateClaimVerdicts  Per-claim verdicts  1
 Step 3  generateArticleVerdict  Overall verdict  1
 Step 4  generateTwoPanelSummary  Format results for UI  0
 Step 5  generateReport  Markdown report generation  0

Budget Controls

 Parameter  Default  Environment Variable
 Max iterations per scope  3  FH_RESEARCH_MAX_ITERATIONS_PER_SCOPE
 Max total iterations  5  FH_RESEARCH_MAX_TOTAL_ITERATIONS
 Max total tokens  100000  FH_RESEARCH_MAX_TOTAL_TOKENS

Shared Modules Used

  • scopes.ts - detectScopes, canonicalizeScopes
  • aggregation.ts - calculateWeightedVerdictAverage, detectClaimContestation, detectHarmPotential
  • claim-decomposition.ts - normalizeClaimText, deriveCandidateClaimTexts
  • budgets.ts - Budget tracking and enforcement
  • provenance-validation.ts - Fact provenance validation

Output Schema

Produces canonical schema compatible with existing Jobs UI:

  • article - Input metadata, thesis, scopes
  • claims - Array of extracted claims with dependencies
  • claimVerdicts - Per-claim verdicts with truthPercentage
  • articleVerdict - Overall verdict
  • facts - Extracted facts from sources
  • sources - Source metadata with reliability scores