Skip to content

Pipeline Visibility Logs

The pipeline details page shows exactly which conversion backend was used for each file and why, along with granular sub-step information for every stage of processing.

Viewing Pipeline Logs

Dashboard

Each completed file in the dashboard table displays a small badge next to its status showing the resolved backend (e.g., marker, claude-vision, cascade:marker+temml). This lets you see at a glance which conversion path was used without clicking into details.

Pipeline Details Page

Click Details on any completed file to open the full pipeline view. The page includes:

  1. Backend Summary Badge - A prominent colored chip at the top showing the resolved backend.
  2. Pipeline Steps Timeline - Each processing stage as a card with duration, cost, and optional sub-steps.
  3. WCAG Compliance Report - Expandable section with evaluated rules, fixes applied, and remaining violations.
  4. Output Preview - The final HTML rendered in an iframe.

Pipeline Steps

StepLabel in UIDescription
routingPipeline RoutingWhich backend was chosen and why
conversionDocument ConversionThe actual file-to-HTML conversion
latex-mathmlLaTeX β†’ MathMLtemml conversion stats (only shown when math tags are found)
image-enhancementImage Analysis & OCRAI-generated alt text for images
image-embeddingImage EmbeddingEmbedding images as data URIs in the HTML
accessibilityAccessibility EnhancementStructural improvements (page headers, heading hierarchy, table headers)
wcag-validationWCAG ValidationFinal WCAG 2.1 AA compliance check and auto-fix

Interpreting the Routing Step

The routing step shows sub-steps that explain why a particular backend was selected.

Auto Mode (parser: 'auto')

When the parser is set to auto, the system runs PDF complexity detection first:

  • Complexity detection - Reports isPureText (true/false) and detected features (e.g., images, tables, math), along with detection time.
  • Routing decision - The reason for the chosen backend:
    • Pure text PDF β†’ marker - Simple text-only documents use the fast/cheap Marker API.
    • Complex PDF (images, tables) β†’ claude-vision - Documents with visual elements use Claude Vision for best quality.
    • No Anthropic key β†’ marker fallback - Falls back to Marker when Claude is unavailable.
    • No Anthropic/Marker keys β†’ mathpix fallback - Last resort fallback.

Cascade Mode (parser: 'cascade')

Cascade mode tries backends in order of cost, escalating if quality is insufficient. Each attempt appears as a separate routing step:

  • Step 1: marker+temml - Cheapest option. Shows the quality score vs. threshold (default 80). If the score passes, this backend is used.
  • Step 3: agentic-vision - Expensive but highest quality. Shows iteration count. Always accepted as the final tier.
  • Fallback: mathpix - Used only when all other cascade steps fail.

Each cascade routing entry includes sub-steps:

  • Backend - Which backend was attempted
  • Quality score - Numeric score vs. threshold (e.g., 85/80)
  • Result - Accepted, Escalating, or Error: <message>

Explicit Parser Selection

When the user explicitly selects a parser (marker, claude-vision, mathpix), the routing step simply records the explicit selection with no complexity detection.

Interpreting the LaTeX β†’ MathML Step

This step only appears when the document contains <math> tags (common in Marker output). It shows:

  • Total math tags found - How many <math> elements were in the HTML.
  • Converted - Successfully converted from raw LaTeX to MathML using temml.
  • Already valid - Tags that already contained proper MathML (no conversion needed).
  • Failed - Tags where temml could not parse the LaTeX (these remain as raw text).

Example: 22 math tags: 20 converted, 1 already valid, 1 failed

A high failure count may indicate the document has unusual LaTeX notation that temml doesn’t support. In that case, cascade mode or claude-vision may produce better results.

Resolved Backend Values

ValueMeaning
markerMarker API (text extraction + temml for math)
claude-visionClaude Vision (page-by-page visual conversion)
mathpixMathpix API (OCR-based conversion)
mammothMammoth.js (DOCX to HTML)
image-passthroughRaw image wrapped in <img> tag
cascade:marker+temmlCascade mode, accepted at the Marker tier
cascade:agentic-visionCascade mode, escalated to Agentic Vision
cascade:mathpixCascade mode, fell back to Mathpix

Backend Badge Colors

ColorBackends
Bluemarker
Purpleclaude-vision
Greenmathpix
Tealmammoth
OrangeAll cascade:* variants
Grayimage-passthrough and unknown