What this type of data looks like
Market data primarily originates from exchange public market data APIs and standardized API interfaces provided by compliant market data service providers. Update cadence is near-real-time or real-time: equity market data pushes the latest trading data every several seconds, while fixed income market data updates at a slower pace. A single standard market data document includes fields such as ticker code, ticker name, latest transaction price, price change amount, price change percentage, trading volume, trading amount, daily highest/lowest price, opening price, and more. Price unit is yuan, trading volume unit is shares or lots, trading amount unit is yuan. Some scenarios include a trading time tag accurate to the second.
Constraints on knowledge base retrieval and recall
The near-real-time or real-time update cadence requires the knowledge base to support high-frequency incremental synchronization, to avoid recalling expired data. The multi-field structured nature of the data requires the retrieval link to support precise filtering by fields such as ticker code and trading time. Without this, unrelated ticker market data will be returned. The standardization of fields and units requires unified unit formatting during data import and retrieval stages, preventing retrieval deviations caused by unit mismatches. The combination of small per-record data size and frequent batch updates means the vector import splitting logic must adapt to single-record or short-period bundled document structures. Excessive splitting will disrupt data relevance.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
recall_top_k | Top 20 results | Market data has moderate per-record information volume; excessive recall will exceed the large model context window limit |
similarity_threshold | 0.75–0.85 | Market data has distinct structured characteristics, with high semantic similarity differentiation; this range can filter low-relevance results |
chunk_size | 800–1200 characters | Adapts to the input length limits of most vector models, while retaining complete information for single-period market data groups |
refresh_interval | 10–30 seconds | Matches the update cadence of real-time market data, ensuring the timeliness of data in the knowledge base |
filter_fields | ["symbol", "trade_time"] | Market data requires precise filtering by ticker code and trading time to avoid returning unrelated tickers or expired data |
rerank_top_k | Top 5 results | User terminal queries typically focus on real-time market data for a small number of core tickers; reranking can focus on the most relevant results |
The parameter values provided on this page are common recommended starting points for configuration setup. Actual values are affected by material form, data volume, and business rules. Specific issues require individual analysis, and it is recommended to test on your own samples before finalizing settings.
Three Common Configuration Errors
- Symptom: In multi-turn query scenarios, retrieval results do not associate historical query ticker information, returning unrelated ticker market data. Cause: The
context_window_enabledparameter is not enabled, and historical query ticker fields are not included in the filtering conditions for the current retrieval. - Symptom: After connecting a custom vector database, valid data cannot be recalled, returning empty results or throwing errors. Cause: The
vector_db_sync_modeparameter is not configured for incremental synchronization mode, or core fields of market data are not mapped to metadata fields of the vector database during import. - Symptom: Garbled characters or field parsing failures occur during retrieval. Cause: The encoding format of imported documents is not specified as UTF-8, or non-standard character sets are not properly handled during upload.
How to Verify Configurations Are Correct
- Execute a single-ticker test query, enter a specific ticker code and trading time, and verify that the
symbolandtrade_timefields of recalled results match the query conditions. - Review vector database synchronization logs, confirm that the incremental synchronization task corresponding to the
refresh_intervalparameter runs at the preset cycle, with no delayed or failed entries. - Trigger a multi-turn query test, enter consecutive queries for different tickers, and confirm that second-round retrieval results automatically associate ticker context from the first round.
- Check retrieval interface return results, confirm that unit fields of recalled results are unified, with no format inconsistencies.
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.