Document parsing acceptance should follow the whole path: original file, extracted text, chunks and indexes, retrieved evidence, and final answer. A successful upload can still contain shifted table columns, repeated headers, missing scanned text, or incorrect citations. Locate each problem at the stage where the discrepancy first appears.
Use this checklist when introducing a document type, replacing a parser, or upgrading related components. Business owners confirm the correct information, implementers inspect processing, and application owners validate answers and citations.
Prepare representative document types
| Type | Sample coverage | Main checks |
|---|---|---|
| Text-based PDF | Multiple columns, headers, footers, and paragraphs across pages | Reading order, repeated noise, and paragraph continuity |
| Scanned PDF and images | Low resolution, rotation, stamps, and mixed languages | OCR completeness, critical numbers, and units |
| Tables | Merged cells, layered headers, and tables across pages | Row-column relationships, inherited headers, and number ownership |
| Word and slide files | Heading levels, lists, captions, and notes | Structure and the relationship between images and explanations |
| Web or linked sources | Navigation, body content, permissions, and updates | Main-text extraction, source access, and update behavior |
Include ordinary documents and known difficult cases. Label critical fields with their page positions and correct values, and preserve original evidence for the evaluation questions. Set thresholds using business risk; amounts, dates, and safety procedures may deserve separate critical-field criteria.
Identify the actual processing path
FastGPT file reading, parsing services, and conversational file input use distinct entry points. Confirm how knowledge-base ingestion and temporary chat files are processed in the deployed version. Record file type, parser, OCR configuration, image handling, and output location.
Upload samples through the intended entry point and compare extracted text with the originals. Check whether image descriptions and table output retain the information needed to answer business questions. Image descriptions generated by a model should be checked against the original image.
The custom enhanced-PDF parser receives a multipart/form-data POST with a file field and returns JSON containing pages and markdown. Configure the service and optional authentication for the deployed version, restart as documented, and enable enhanced PDF parsing at ingestion. Use the same file to compare the parser response with the content actually ingested.
Validate four layers
- Parsing: compare critical fields, reading order, table relationships, and missing content against the source, recording the affected page.
- Chunking: check that headings remain connected to content, cross-page material stays coherent, and headers or units accompany data. Sample unusually short, long, or repeated chunks.
- Retrieval: use fixed questions to inspect whether the first k results contain sufficient supporting evidence. Record empty retrieval and confusion between similar documents.
- Answering: check consistency with evidence, citation targets, and behavior when the material is insufficient.
Each question should be traceable through the original file, chunk, retrieved result, and answer. Fix the earliest stage that introduced the discrepancy, then repeat the affected checks.
Include abnormal inputs
Exercise password-protected, corrupted, oversized, timed-out, and unavailable-source inputs. Record the user-visible status, processing-queue result, and available recovery action. For batch ingestion, reconcile successful, failed, duplicate, and pending items so every input has an outcome.
Before resubmitting a file, inspect existing records to avoid duplicates that alter retrieval. After changing parser settings, process the same samples again while keeping the generation model, retrieval parameters, and questions stable.
Retain a reusable acceptance record
Record the file identifier, type, parser configuration, error location, affected question, owner, and repeat-test result. Add critical errors and boundary samples to the fixed checks used for later upgrades. Continue adding real failure cases during the observation period so acceptance reflects actual business inputs.
Related guides
- Enterprise Document Parsing Tool Selection Validation Guide
- FastGPT Knowledge Base Lifecycle and Ownership Guide