What the Data for This Category Looks Like
Precast concrete components include PC components, pipe piles, bridge components, and municipal products. Data sources include BIM or detailed design documents, steel bar and concrete batch ledgers, mold table and curing kiln operation records, laboratory test reports, and QR code traceability data. Daily documents include component detailed drawings, steel bar cutting lists, mold inspection records, pouring and vibration records, curing records, and demolding and factory inspection reports. Data update frequency follows per individual component, per batch, or per kiln. Document fields include component ID, steel bar specification, batch number, curing parameters, strength test values, and more. Single-document length varies widely, ranging from short-text test item descriptions to long-text detailed drawing annotations.
How These Characteristics Impact Vector Model and Indexing Workflows
The high-frequency update nature of this category requires indexes to support incremental updates. This avoids resource consumption and delays caused by full reindexing. The mixed scenario of multiple document types requires chunking strategies adapted to different text lengths. This prevents over-splitting long documents or losing key context for short documents. Multi-source data requires unified field mapping rules. This ensures consistency of core identifiers such as component IDs and batch numbers, preventing matching deviations during retrieval. Additionally, the business characteristic where errors carry severe consequences requires high precision in retrieval results. Strict control of recall relevance thresholds is needed to prevent component scrap or project delays caused by mismatches.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-text length of steel bar cutting lists and curing records. Prevents over-splitting chunks and losing key identifiers like component IDs |
| Overlap Length | 100–150 characters | Connects core fields of adjacent chunks. Ensures complete information such as batch numbers and component IDs across chunks |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
defaultToken | 256 token | Provides a baseline embedding length for short-text mold inspection records. Avoids excessive truncation of short texts |
| Similarity Threshold | 0.75–0.85 | Filters low-match irrelevant documents. Prevents component installation or quality issues caused by mismatches |
| Single-Search Limit | Top 10 results | Covers multiple document types corresponding to a single component. Ensures retrieval of all associated content such as cutting lists and inspection records |
This configuration applies to FastGPT v4.17.0
Three Common Configuration Errors
- Phenomenon: The number of retrieval results is far fewer than expected. The interface displays "No matching results". Cause: Chunk length is set too large. Complete detailed drawing text is split into too few chunks, which cannot match precise keyword segments.
- Phenomenon: Custom index configurations do not take effect. Target documents cannot be recalled during testing. Cause: Index mapping is not established based on core fields such as component IDs and batch numbers. This prevents correct association between data and retrieval keywords.
- Phenomenon: Retrieving documents for a batch of pipe piles recalls curing records for other components of the same category. Cause: The similarity threshold is set too low. Low-relevance text fragments are misjudged as matching items.
How to Verify Correct Configuration
- Upload a single steel bar cutting list for a precast concrete component. Execute a retrieval request for "verify steel bar positions". Check if mold inspection records for the same component are recalled.
- View the index update log. Confirm that incremental indexing is automatically triggered after documents for each batch of components are uploaded. Full reindexing is not triggered.
- Adjust the similarity threshold. Observe changes in the matching degree of retrieval results. Verify that the threshold configuration aligns with the document characteristics of the category.
- Test using the component ID as the retrieval keyword. Confirm that all associated documents for the corresponding batch are accurately recalled.
The values provided are common starting points and should be measured against the reader's own samples.