What this category of data looks like
This category of data originates from BIM detailed design at precast concrete component factories, steel and concrete batch ledgers, mold table and curing kiln operation logs, laboratory test reports, and QR code traceability systems. Daily documents include component detailed drawings, steel bar cutting lists, mold inspection records, pouring and vibration logs, curing records, and demolding factory inspection reports. Fields cover component number, steel bar diameter, embedded part position, pouring strength, curing duration, and more. Update frequency adjusts in real time based on individual components, each material batch, or each kiln curing batch.
Constraints on the Document Parsing and Chunking Link
Document types span multiple formats including engineering drawings, ledger records, and test reports. Some documents are PDFs exported from CAD, so vector text extraction must be supported. Update frequency adjusts per component or batch, so chunking must split content by component or batch to prevent text overlap across units. Many fields are closely linked, so chunking must retain associated identifiers such as component numbers to enable accurate list content verification during bid quotation. Parsing large detailed drawings and multi-page records takes significant time, so timeout interruptions must be avoided.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the text length of precast component detailed drawings and steel bar cutting lists, retains associated fields such as component numbers and steel bar parameters |
chunkSplitter | Split by component ID + paragraph | Matches the independent document unit of each component, avoids text overlap across components |
customPdfParse | Enable vector text extraction mode | Adapts to CAD-exported PDF documents from BIM, retains precise coordinate and size information in drawings |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Covers the parsing duration of large detailed drawings and multi-page curing records, avoids timeout interruptions |
chunkSettingMode | Manual configuration | Adjust chunking rules for different document types such as cutting lists and curing records, adapts to specific scenario requirements |
Three Common Misconfigurations
- The
detail: trueparameter during external calls cannot be parsed synchronously with knowledge base call parameters and streaming return results, leading to empty return fields. This occurs when external call parameters are not uniformly bound to document parsing logic, causing context transfer interruptions. - Importing Excel format files fails, and CSV files only recognize two columns of data. This happens when custom file parsing rules are not configured. FastGPT’s default CSV parsing only supports a fixed number of columns, and does not adapt to multi-field cutting lists and batch data for precast components.
- Connection failures appear after deploying marker-pdf, and access using 127.0.0.1 returns an error. This stems from network isolation between Docker containers. The parsing service port is not mapped to the network segment accessible to FastGPT, resulting in connection timeouts.
How to Verify Correct Configuration
- Upload a single PDF of a precast component detailed drawing. Confirm the parsed text blocks include core fields such as component number, steel bar diameter, and embedded part position, with no obvious text truncation or overlap.
- Initiate an external call with the
detail: trueparameter. Verify the return result includes both knowledge base call parameters and document chunk content from streaming parsing, with no missing fields. - Upload a multi-column CSV cutting list file. Confirm all columns of steel bar specifications, quantities, and weight information are extracted after parsing, without being restricted by the default two-column limit.
- Check the connection status of the marker-pdf service in the FastGPT v4.17.0 background management interface. Confirm no abnormal error reports are present.
The values listed here are common starting points and should be evaluated against the reader’s own samples.