Knowledge Base Retrieval and Recall for Securities Research Report Queries

Data sources for securities research reports primarily include public reports from licensed securities research institutions, and research report

What the data for this category looks like

Data sources for securities research reports primarily include public reports from licensed securities research institutions, and research report content from compliant authorized financial data terminals. Updates are pushed in batches after each trading day’s close. Reports related to major events are updated in real time.

Document structure typically includes a cover page (containing report title, issuing institution, analyst, publish date), core summary, industry analysis, individual stock ratings, profit forecasts, and risk warnings.

Core fields include: unique report identifier, issuing institution name, analyst name, investment rating, target price (unit: RMB yuan), profit forecast value (unit: 100 million yuan). Most documents are in PDF format, with some exportable to structured text formats.

Constraints on Knowledge Base Retrieval and Recall

High-frequency report updates require the retrieval pipeline to support incremental index updates, to avoid performance losses from full index rebuilding. The presence of structured fields requires configuring structured retrieval branches, to distinguish between text semantic recall and numerical condition filtering, and improve matching accuracy.

Individual research reports can be lengthy, with some reaching tens of thousands of characters. Adjusted chunking strategies are needed to preserve the integrity of logical units, while controlling single chunk length to fit context windows.

In addition, strict compliance requirements apply. Recalled content must retain issuing institution and publish date metadata, to avoid unauthorized use of non-compliant content.

Configuration Parameters

Configuration ItemRecommended ValueRationale
RECALL_CHUNK_SIZE800–1200 charactersLogical units for individual securities research report chunks mostly fall within this range, avoiding splitting that disrupts the integrity of core content such as industry analysis and ratings
SIMILARITY_THRESHOLD0.75–0.85Research report content is highly specialized. This range filters low-relevance fuzzy matching results while retaining enough candidate content for reranking
MAX_RECALL_NUMTop 10 resultsSingle research report retrieval needs typically focus on the latest reports in the same industry and with the same rating. Too many recalled results increase context window pressure
PARSE_FILE_TIMEOUT_SECONDS300 secondsParsing long single PDF research reports takes significant time. This setting prevents upload failures caused by parsing timeouts
ENABLE_STRUCTURED_RECALLEnabledResearch reports include structured fields such as target price and profit forecasts. Enabling this setting supports numerical range retrieval and improves accuracy
INDEX_UPDATE_CRON0 2 * * *Research report updates are concentrated after daily market close. Incremental index updates at 2 AM ensure content timeliness for the next day’s retrieval

The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing settings.

Three Common Configuration Mistakes

  • Phenomenon: The number of retrieval results returned is far lower than expected. Cause: SIMILARITY_THRESHOLD is set too high, filtering out a large number of semantically relevant research report passages.
  • Phenomenon: Retrieval results include unauthorized non-official research report content. Cause: External network retrieval configuration is not disabled, causing the system to automatically crawl non-compliant research reports from public networks.
  • Phenomenon: Research report uploads fail during parsing, returning a 408 Request Timeout error. Cause: PARSE_FILE_TIMEOUT_SECONDS is set too short, terminating parsing before long PDF format research reports are completed.

How to Confirm Configuration is Correct

  • Upload 1 to 2 official research reports with different publish dates. Verify that metadata including issuing institution, publish date, and target price is retained after parsing, to confirm structured parsing configuration is active.
  • Submit a retrieval request that includes specific investment ratings and target price ranges. Verify that matching research report content is returned, to confirm structured retrieval configuration is working properly.
  • Review index update logs. Confirm that incremental indexes run at the preset schedule with no failed entries, to confirm update configuration is active.
  • Adjust SIMILARITY_THRESHOLD to different ranges, compare the number of retrieved results, to confirm that the threshold configuration properly impacts recall results.

Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.