What Data for This Category Looks Like
Data for this category typically comes from ERP systems, supplier shipping documents, logistics tracking information, and device manufacturer backends. Daily generated documents include purchase orders, delivery acceptance slips, serial number lists, and invoice materials. A single purchase order links multiple devices, with fields such as device model, serial number, and delivery time. Data updates follow per-order or weekly cycles. Individual records are generated for the full device delivery process of a specific integration project. Core fields include project number, device model, supplier information, and acceptance status, among others.
Constraints Imposed by These Characteristics on Reranking and Hybrid Retrieval
Records for this category are built around single-project device delivery. Individual records have many fields, and link to strongly matched fields such as project number and serial number. Retrieval requires precise association with project context. Data update frequency is uneven. A single retrieval may need to cover multiple documents from the past 7 days up to the full project cycle. Document lengths vary widely: some are short serial number lists with tens of characters, while others include hundreds of characters of detailed purchase order content. Hybrid retrieval must balance structured field matching and unstructured text. The reranking step must prioritize matching core identifiers from user queries, such as device model and serial number. It must also prevent irrelevant cross-project data from interfering with retrieval results, and adapt relevance calculation logic for documents of different lengths.
Configuration Settings
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches strong identifying fields like device model and serial number for this category. A too-low value will allow irrelevant data to be included, while a too-high value will miss relevant project documents |
limit | Top 10–15 results | Adapts to the number of device documents linked to a single project. Prevents excessive retrieval results from exceeding the context window, while covering all relevant delivery materials |
searchMode | hybrid | Combines structured field matching and vector retrieval, adapting to the multi-source, multi-structure document characteristics of this category |
usingReRank | true | Reranks retrieved results by relevance, prioritizing matching core user query requirements such as device model and delivery time |
| Rerank Trigger Threshold | Trigger when ≥8 results are recalled | Filters low-relevance content via reranking when there are many recalled results, adapting to the wide variation in document length for this category |
The above configurations are designed for the standard retrieval pipeline of FastGPT v4.17.0.
Three Common Misconfigurations
- Phenomenon: The reranking model works during independent testing, but the result reranking flag is marked
falseduring formal retrieval. Cause: FastGPT’s retrieval pipeline did not correctly pass the reranking model call parameters, or the API key configuration is incorrect. - Phenomenon: When the number of retrieval results does not reach the threshold, the reranking model does not take effect, and result sorting is only based on vector similarity. Cause: The reranking model’s recalled result count threshold is not configured, or the threshold is set higher than the actual number of recalled results.
- Phenomenon: Hybrid retrieval only recalls documents matched by structured fields, and does not include unstructured delivery acceptance notes. Cause:
searchModeis configured for structured retrieval only, and hybrid retrieval mode is not enabled.
How to Verify Proper Configuration
- Access the FastGPT knowledge base test page, submit a query containing device model and serial number, and review whether the sorting of retrieval results aligns with expected relevance.
- Review retrieval logs to confirm that the
usingReRankfield istrueand that a call record for the reranking model exists. - Adjust the
similaritythreshold, observe changes in the number of retrieved results, and verify that the threshold configuration is active. - Switch
searchModeto pure vector retrieval, compare result differences with hybrid retrieval, and confirm that the hybrid mode configuration is correct.
The values provided are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.