What the data for this category looks like
Daily documents for this category include keyword lists, negative keyword tables, bid records, and search term reports. Data is sourced from search ad backends, bidding ranking data, search term reports, and other channels, with a daily or weekly update frequency. Most documents use structured formats, containing fields such as keyword text, match type, bid amount, and negative keyword rules. Some reports store multi-dimensional data in table form. Unlike creative documents for feed ads, these materials focus on user-initiated search keywords and their associated rules.
What constraints these characteristics impose on document parsing and chunking
Most documents are structured tables or plain text with defined fields. Compliance reviews require retaining the association between keywords and their corresponding rules. Chunks cannot be too small, as this will split associated information. They also cannot be too large, as this will prevent individual review records from being located. Daily updated batch documents also require an efficient parsing process to avoid excessive resource consumption. Different document formats (excel, pdf, txt) need dedicated field extraction logic. Without this, data loss or format errors may occur, disrupting subsequent compliance checks.
How to configure settings
The following configuration items are based on FastGPT v4.17.0, and are adapted to the document parsing and chunking requirements for search ad keyword management.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the keyword plus context length of search ad documents, avoiding splitting the association between negative keywords and their corresponding search terms |
chunkSplitter | Split by paragraph + keyword row | Matches the structured row format of negative keyword tables and search term reports, retaining complete information required for individual compliance reviews |
customPdfParse | Enable structured extraction mode | Adapts to table-based PDF documents such as search term reports, restores the original field structure and prevents content loss |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Addresses batch daily updated search term report files, avoiding parsing timeouts for large documents |
chunkSettingMode | Manual configuration | Adjust chunking rules for different document types (excel bid sheets, txt keyword lists) to meet the field association requirements of compliance reviews |
Three common configuration errors
- Scenario: After uploading excel-format bid records, the parsing result only displays unassociated plain text, and the bid amount and match type fields linked to cells are lost. Cause: The excel structured parsing mode of
customPdfParseis not enabled. The default parsing process only extracts unassociated plain text content. - Scenario: After uploading a single search term report larger than 100MB, a parsing task timeout error is triggered. Cause: The
PARSE_FILE_TIMEOUT_SECONDSconfiguration value is set too low, and the timeout threshold is not adjusted for large-volume batch documents. - Scenario: When replacing a third-party document parsing tool, a large number of missing fields appear in the parsing result. Cause: The adaptation parameters of
customPdfParseare not correctly configured, and the tool is directly replaced without completing integration with the FastGPT parsing workflow.
How to verify a successful configuration
- Upload a single keyword list document, and confirm that parsed chunks include complete keyword and match type combinations with no missing fields.
- Review parsing task logs to confirm no timeout errors matching the
PARSE_FILE_TIMEOUT_SECONDSthreshold are triggered. - Upload a PDF-format search term report, and verify that all fields from the embedded table are extracted with no content loss.
- Adjust the
chunkSizevalue, compare chunking results across different configurations, and confirm that keywords and their associated rules are not split across multiple chunks.
The values provided are common starting points and should be evaluated against samples relevant to the specific deployment.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.