What this category of data looks like
Data for this category originates from ERP orders and inventory, bills of materials (BOMs) and process routes, equipment capacity, staff shifts, work-in-progress items, and supplier deliveries. Daily processed documents include sales orders, master production schedules, material requirements plans, weekly scheduling tables, work orders, material shortage lists, and delivery change notices. Data is updated daily, weekly, or triggered by orders or abnormal events. Data fields include order numbers, material codes, capacity values, shift durations, work-in-progress quantities, and more, with corresponding units such as pieces, hours, days, and quantities.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Multi-source, heterogeneous data sources require retrieval to support cross-data-source association matching, to prevent retrieving unrelated, scattered data. The high-frequency update feature requires the recall mechanism to support incremental indexing or real-time synchronization, ensuring retrieval results match the latest scheduling and order statuses. Diverse document structures and fields require retrieval to divide scope by business collections, while supporting precise matching of structured fields. Differences in priority across scenarios require recall results to be reordered according to business rules, prioritizing urgent work orders or delivery change requests.
How to set configurations
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the field association accuracy of production planning and scheduling documents, to avoid retrieving unrelated equipment logs or quality inspection records |
limit | Top 8 entries | Adapts the single retrieval context length for shop scheduling scenarios, to avoid exceeding the large model's token limit |
searchMode | embedding | Performs precise recall for structured material, capacity, and scheduling data, replacing fuzzy full-text retrieval |
usingReRank | Enabled | Reorders recall results by work order priority and delivery urgency, aligning with the business logic of supplier and process quality audits |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonym associations related to scheduling, such as associating "work order" with "production task order" to cover more synonymous expressions |
collectionIds | Business-divided scheduling and supplier audit collections | Limits the retrieval scope, to avoid mixing unrelated cross-category documents |
Three common mistakes
- Calling the retrieval API without specifying the
collectionIdsparameter results in an empty JSON response. The retrieval scope covers the entire platform knowledge base, and no production planning-related business collections are matched. - Adding or deleting data to or from a specified collection via the API without carrying the correct structured field mapping causes data to fail proper indexing, and unable to be matched during retrieval.
- Directly using raw HTTP response business data as knowledge base references, without converting it to the structured format required by FastGPT, causes references to be unrecognizable by the retrieval engine, resulting in no matching results after retrieval.
How to confirm the configuration is correct
- A specific scheduling scenario question, such as "Machining work order schedule for the second week of May 2024", is used to verify whether retrieval results include work order data for the corresponding time period.
- Retrieval logs are reviewed to confirm that the matching value of the
similarityparameter falls within the preset 0.75–0.85 range, with no overly low or high matching results. - When
usingReRankis enabled, verification is performed to confirm that high-priority urgent work orders appear at the top of retrieval results. - The API is called to retrieve the collection specified by
collectionIds, and confirmation is made that only document data within that collection is returned, with no unrelated cross-collection content.
The values provided are common starting points and should be validated using local sample data.