Data Overview for This Category
The data for this category primarily comes from search ad backends, search term reports, bidding ranking data, and landing page data. Most updates follow a daily or weekly schedule. Daily produced documents include keyword lists, negative keyword tables, bid records, and search term reports. Fields covered include keyword text, match type, single bid amount, click volume, conversion cost, and more. Units include yuan, clicks, entries, and others. The number of entries per document fluctuates based on the size of the ad account.
Constraints on Knowledge Base Retrieval and Recall
The daily or weekly update rhythm of the data requires knowledge base synchronization tasks to match this frequency. Data lag is prevented from impacting the timeliness of keyword adjustments. Documents include structured bid and click data, plus semi-structured negative keyword and keyword text data. The retrieval process must support multi-field associated recall. It must also filter expired historical data entries. The number of entries per document changes with account scale. The system must adapt to retrieval loads of different magnitudes. This avoids recall timeouts or reduced accuracy caused by excessive data volume. Retrieval for ad scenarios must prioritize associating the currently optimized keyword groups. This ensures recall results are directly relevant to business goals.
Configuration Settings
This documentation is written for FastGPT v4.17.0. The following configuration items are optimized for the search ad keyword management scenario:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the relevance of negative keywords and keywords, filters low-correlation results, and avoids invalid recall |
limit | Top 10–15 entries | Adapts to the entry scale of a single ad document, controls context length, and avoids exceeding model input limits |
searchMode | Keyword matching + semantic recall | Covers structured keyword fields and semi-structured search term report content, balances precise and generalized retrieval |
usingReRank | Enabled | Re-ranks initial recall results by business relevance, prioritizes returning keywords or negative keywords with high conversion potential |
datasetSearchUsingExtensionQuery | Enabled | Expands relevant search terms based on the original question, covers more potential keyword expansion scenarios |
collectionIds | Bind exclusive ad knowledge base collections | Isolates business data, avoids interference from documents of other categories in retrieval results |
Three Common Configuration Mistakes
- Retrieval results include documents from non-ad categories. This occurs when
collectionIdsis not configured correctly. The retrieval scope covers unisolated other knowledge base collections. - Some PDF documents fail to parse. This happens when document text exists as scanned image format. OCR preprocessing is not enabled, and FastGPT supports only directly text-extractable PDF files by default.
- Retrieval responses time out, or the number of results is abnormal. This stems from two issues. The
limitparameter is not set appropriately, leading to excessive recalled entries that exceed the model's input context limit. Or the synchronization task does not match the data update frequency, resulting in simultaneous loading of large data volumes.
How to Verify Successful Configuration
- Run a retrieval targeting a negative keyword table. Check that returned results only include entries related to negative keywords for this category, with no unrelated documents included.
- Upload a scanned PDF document. Confirm that the parsing status shows no errors. Verify that OCR preprocessing is working correctly.
- Adjust the
similarityparameter to 0.8. Run a retrieval for a keyword expansion-related question. Compare the relevance of results before and after the adjustment to confirm the parameter takes effect. - Check the knowledge base synchronization logs. Confirm that the synchronization frequency matches the daily or weekly update rhythm of the business. No lag or duplicate synchronization occurs.
The values listed here are common starting points. They should be evaluated against relevant 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.