What this category of data looks like
Data for this category typically comes from customer interviews, bidding documents, vendor product materials, and historical project plans. Daily used documents include technical proposal documents, customer requirement survey forms, quotation lists, and demo scripts. Document structures cover requirement items, technical routes, quotation details, and demo steps. Some fields include identifiers such as project number, requirement priority, and quotation unit. Update frequency is per order or weekly. Document lengths vary widely, from hundreds of characters for requirement items to tens of thousands of characters for complete proposal documents.
Constraints on knowledge base retrieval and recall
Dispersed data sources create significant differences in text expression styles. Colloquial content from customer interviews and formal wording in bidding documents can cause semantic mismatches. Retrieval models must adapt to matching multi-style text. High update frequency and order-based updates require the retrieval system to support incremental knowledge base synchronization. This avoids extra time spent on full re-indexing. Wide variation in document length means short-text requirement items and long-text proposal documents need tailored chunking and recall strategies. The core goal of pre-sales proposals is to match the current project’s boundaries and technical routes. This requires avoiding irrelevant historical project documents and strictly limiting retrieval scope.
Configuration settings under FastGPT v4.17.0
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Adapt to the technical terminology and multi-style text of pre-sales proposals, filter low-relevance recall results |
limit | Top 8–12 entries | Balance recall coverage and context window usage, avoid excessive long text exceeding processing limits |
usingReRank | Enabled | Perform secondary ranking on initial recall results, filter historical proposals that are semantically similar but irrelevant to the current project |
datasetSearchUsingExtensionQuery | Enabled | Expand the semantic scope of retrieval terms, adapt to retrieval scenarios where wording in customer requirement survey forms does not directly match |
collectionIds | Bind dedicated knowledge base collection for pre-sales proposals | Restrict retrieval scope, avoid retrieving documents from other categories, improve accuracy |
Three common mistakes
- Retrieval returns non-target documents. For example, when querying a customer’s requirement items, the system returns proposals from other projects. This occurs when
collectionIdsis not configured to restrict retrieval scope, so retrieval covers the full knowledge base. - Garbled characters appear in the pre-sales proposal knowledge base after online deployment, while local deployment works normally. This happens when the online environment’s character encoding is not set to UTF-8, so special characters in vendor product materials cannot be parsed correctly.
- The knowledge base search result module is displayed on the interface and cannot be hidden. This occurs when the front-end knowledge base search display switch is not turned off, or the retrieval entry is not hidden in the system configuration.
How to confirm successful configuration
- Check the bound ID of the knowledge base collection, confirm it only includes pre-sales proposal-related technical proposal documents, customer requirement survey forms, quotation lists, and demo scripts.
- Submit a test query, enter a specific entry from a customer requirement survey form, verify that recall results only include documents directly related to that entry.
- View the similarity scores of initial recall results, confirm they fall within the preset threshold interval of
similarity. - Upload a new quotation list, verify that the knowledge base completes incremental updates within 10 minutes and supports retrieval.
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.