Document Parsing and Chunking for Advertising and Marketing

What data for this category looks like

Data sources for advertising and marketing intelligent due diligence reports include placement ledgers submitted by advertising agencies, media procurement contracts from brand parties, and placement effect reports from third-party monitoring agencies. Update rhythm adjusts with individual due diligence project cycles, and core documents are submitted in batches per project milestones. Document structures mostly include structured tables, long-text analysis paragraphs, embedded effect data charts. Some documents are PDFs converted from scanned files. Core fields include placement date, media channel, placement amount, impressions, conversions; corresponding units are day, channel name, yuan, times, times.

Constraints on document parsing and chunking

High proportions of structured tables require parsing components to retain row-column hierarchical associations, and avoid splitting cells which would misplace placement detail data. Large proportions of long-text analysis paragraphs require retaining logical associations within paragraphs during chunking, and avoid cutting cross-topic effect interpretation content. Field units are scattered and require clear labeling. During parsing, extract fields and their corresponding units simultaneously to prevent unit confusion in subsequent retrieval. Some documents are PDFs converted from scanned files, with text layer offset issues. Parsing processes adapted to non-standard text layouts are needed. Update rhythm adjusts flexibly with projects. Parsing nodes must support on-demand adaptation to varying document volumes.

How to set configurations

Configuration Item	Recommended Value	Rationale
`parse_table_mode`	`preserve_structure`	Advertising and marketing due diligence reports contain a large number of placement detail tables. Retaining structure prevents breakage of data associations
`chunk_max_length`	`800–1200 characters`	Analysis paragraphs in advertising and marketing due diligence reports are mostly coherent effect interpretations. This length retains the integrity of single-topic analysis
`parse_file_timeout_seconds`	`300 seconds`	A single due diligence report may include multiple attachments. A longer timeout covers parsing processes for large-volume documents
`extract_metadata_fields`	`["投放日期", "媒介渠道", "投放金额", "曝光量"]`	Core retrieval fields for advertising and marketing due diligence reports are listed above. Structured extraction in advance is supported
`enable_ocr_for_pdf`	`true`	Some third-party monitoring reports are in scanned file format. OCR restores complete text content
`chunk_overlap`	`100–150 characters`	Cross-paragraph logical associations in advertising and marketing reports are strong. Overlap retains context coherence

The parameter values provided on this page are common starting points for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require specific analysis. It is recommended to test on internal samples before finalizing.

Three common mistakes

Symptom: After front-end packaging and deployment to the server, uploading an advertising and marketing due diligence report returns a 404 error, and the document parsing node cannot obtain file content. Cause: The server environment’s file upload path is not configured with cross-domain access permissions, or the static resource directory is not mapped to the correct file storage path.
Symptom: When deploying Qwen3-14B using Vllm 0.10, preset fields cannot be extracted from PDF parsing results. Cause: Vllm 0.10 has varying adaptation for structured output parsing of some large language models, leading to failure of the field extraction logic.
Symptom: Parsed document chunks lose table row-column associations, and complete placement detail data cannot be matched during retrieval. Cause: parse_table_mode is not configured as preserve_structure, and the default parsing mode splits the table structure.

How to confirm configurations are correctly set

Upload a standard advertising and marketing due diligence report docx document, check the parsed chunk content to confirm the table structure is fully retained.
Check the parsing log to confirm that the extracted metadata fields match the configured extract_metadata_fields list.
Test documents in different formats (regular PDF, scanned PDF) to confirm that parsing results comply with the enable_ocr_for_pdf configuration requirements.
Adjust the chunk_max_length parameter, then upload a test document to verify that the chunk length falls within the preset range.

Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.