What data for this category looks like
Data sources include architectural glass product specifications, third-party test reports, industry standard documents, and marketing promotion material packs. Update frequency adjusts alongside new product production and industry standard revisions, with no fixed cycle. Most documents are PDF or Word files containing parameter tables. Each document corresponds to a single glass product type. Fields include glass type, thickness, compressive strength, applicable scenarios, and certification number. Common units are millimeters (mm), megapascals (MPa), and kilograms per square meter (kg/㎡).
What constraints these characteristics impose on knowledge base retrieval and recall
The single-document-per-glass-product-type structure leads to fine-grained documents in the knowledge base. Retrieval must prioritize matching precise product type keywords. Most documents include parameter table structures. When chunking content, retain the relational links between fields in tables to avoid losing corresponding logic after parameter splitting. Physical parameter fields mean user queries often include specific numerical requirements. Retrieval logic must prioritize matching parameter similarity. The lack of a fixed update cycle requires configuring incremental update mechanisms to replace outdated parameter documents promptly. The presence of marketing material packs means retrieval results must cover both product parameters and scenario adaptation instructions to meet customer acquisition needs.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Glass product documents often include multi-page parameter tables. 600 seconds covers parsing time for most large-sized documents |
Chunk Length | 800–1200 characters | Adapts to the combined structure of parameter tables and explanatory text in glass documents, avoiding loss of field association after splitting |
Recall Count | Top 6 results | Balances multi-product coverage required for glass marketing customer acquisition and information reading load. 6 results meet requirements for most scenarios |
Similarity Threshold | 0.72–0.78 | Adapts to the precise matching needs of glass parameters, preventing irrelevant content from being recalled if set too low, and qualified materials being missed if set too high |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Covers upload requirements for multi-page high-resolution scanned glass test reports, preventing upload failures for large documents |
Reranked Return Count | Top 3 results | Prioritizes displaying the most matching product parameters and scenario cases, reducing user filtering costs |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require on-site analysis. Testing against internal test samples is recommended before finalizing settings.
Three Common Configuration Mistakes
- Symptom: Parsing progress stalls or timeout errors are returned when uploading large glass documents. Cause: The
PARSE_FILE_TIMEOUT_SECONDSparameter was not adjusted. The default timeout duration is insufficient for parsing multi-page parameter tables. - Symptom: Glass parameters and applicable scenarios in retrieval results are displayed separately, with no corresponding association. Cause: Table structures were not retained during chunking. Parameter tables were split into independent segments, losing the corresponding relationship between fields.
- Symptom: Queries unrelated to the knowledge base still return glass product content from the knowledge base, and the preset fallback reply is not triggered. Cause: No fallback rule was configured for the similarity threshold. When the matching degree is below the threshold, the specified reply branch is not activated.
How to Confirm Configurations Are Correctly Deployed
- Upload the largest-sized single glass document, verify parsing progress and duration meet expectations. Adjust
PARSE_FILE_TIMEOUT_SECONDSto a suitable duration. - Randomly select one glass product document, verify chunked content retains the complete structure of parameter tables. Adjust chunk length to an appropriate range.
- Submit a test query unrelated to the knowledge base, verify the preset fallback reply is triggered. Confirm the similarity threshold and fallback logic configuration is active.
- Submit a test query containing specific glass parameters, verify the number and matching degree of recall results. Adjust recall count and similarity threshold to meet business requirements.
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.