What the Data for This Category Looks Like
Data sources for ticket auto-classification mainly include government hotline tickets, public service knowledge base documents, and user inquiry records. Data updates follow two rhythms: real-time new ticket creation, and daily batch synchronization of knowledge base files. A single ticket data entry includes structured fields such as appeal text, classification tags, and processing workflows. Knowledge base documents include policy files, frequently asked questions, and batch-exported ticket statistics tables. Some attachments include PDF and image-format scanned paper ticket copies. Data fields are mostly text and structured tables, with units measured in entries, characters, and MB.
What Constraints Do These Characteristics Impose on the Document Parsing and Chunking Link
The appeal text in hotline tickets varies widely in length, from simple inquiries of tens of characters to complex questions of hundreds of characters. Chunking must retain semantic integrity to ensure classification accuracy. Bulk-imported ticket tables contain multiple columns of structured fields. Chunking must avoid truncating column-related information, otherwise classification tags will be lost. Real-time updated ticket data requires the parsing process to have low latency. Excessive parsing timeout will affect ticket circulation efficiency. PDF attachments containing images cannot directly extract valid text. OCR parsing capability is required, otherwise key ticket information will be lost.
How to Set Configurations
| Config Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Ticket appeal text mostly falls within the hundreds-of-characters range. This chunk length retains complete semantics and prevents classification logic from being truncated |
chunkSplitter | Semantic chunking | Ticket classification relies on semantic associations of appeals. Semantic chunking prevents classification bias caused by rigid splitting |
PARSE_FILE_TIMEOUT_SECONDS | 120–180 seconds | When deploying the Qwen14B 32K model locally, parsing large ticket tables or PDF attachments takes longer. This threshold prevents timeout failures |
UPLOAD_FILE_MAX_SIZE | 500 MB | Hotline knowledge bases often contain batch-exported monthly ticket tables and multi-page policy PDF collections. This upper limit covers conventional import requirements |
customPdfParse | Enabled | Ticket attachments often include scanned PDF files. Enabling custom parsing triggers the OCR process and extracts text content from images |
chunkSettingMode | Manual configuration | Ticket data structure is fixed. Manual configuration allows precise control over chunking logic to adapt to the input requirements of classification models |
Three Common Mistakes
- Symptom: Uploading a PDF ticket attachment containing images results in no text extracted from the images. Cause: The
customPdfParseconfiguration is not enabled, so the OCR parsing process is not triggered, leading to loss of image information. - Symptom: Uploading a table dataset results in some data fields being empty, making training data incomplete. Cause: Chunk length is set too small, truncating column-related information and preventing structured data from being fully retained.
- Symptom: Uploading a file results in a request failure after parsing takes more than 2 minutes, and restarting the server does not fix the issue. Cause:
PARSE_FILE_TIMEOUT_SECONDSis not adjusted to match local large model computing power. The default threshold cannot cover the parsing duration of large files.
How to Confirm the Configuration Is Complete
- Upload a single 100 MB monthly ticket export table, check that there are no timeout errors during parsing and that parsing duration meets expectations.
- Upload a ticket PDF attachment containing images, confirm that the parsed text content includes handwritten or printed text from the images.
- Export the current FastGPT v4.17.0 knowledge base configuration file, compare parameters such as
chunkSizeandchunkSplitterwith the new server to ensure complete configuration consistency. - Submit 10 test tickets of different types, check that the chunk association logic of classification results matches the preset configuration.
The values provided are common starting points and should be measured against relevant local samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.