Vector Models and Indexing for Heating Industry Marketing

Data Profile

Data sources include official WeChat public account posts from heating enterprises, electronic versions of offline promotional materials, compiled frequently asked questions from user service hotlines, and copy for heating season exclusive marketing campaigns. Updates occur in concentrated batches 1–2 months before the heating season starts, 1–2 times per month during regular operations, and new content is added on demand during marketing campaigns. Most documents are long texts, containing fields such as event rules, service hours, charging standards, and service coverage areas. Involved units include temperature (℃), payment amount (yuan), and service coverage scope (streets/communities).

Key Constraints for Vector Models and Indexing

The long text structure of heating marketing content requires vector models and indexing to balance segmentation integrity and semantic relevance, and avoid splitting core information such as event rules and area descriptions. Clear units in fields such as temperature and payment amount require indexes to retain numerical association features, preventing semantic deviation from generalized encoding. The centralized update rhythm and on-demand content addition require indexes to support incremental updates, avoiding full retraining to reduce resource consumption. Timely marketing content also requires indexes to support configurable expiration times, preventing recall of outdated campaign information.

Recommended Configuration Settings

Configuration Item	Recommended Value	Rationale
`chunk_size`	800–1200 characters	Heating marketing content contains long segments such as event rules and area descriptions. This range preserves semantic integrity and avoids splitting core information
`chunk_overlap`	100–150 characters	Content such as service areas and charging standards may span segments. The overlapping interval preserves contextual association and avoids semantic breaks
`RECALL_TOP_K`	Top 6–8 results	User queries in heating marketing scenarios usually focus on a single scenario. Too many recalls will introduce information from unrelated areas or campaigns
`SIMILARITY_THRESHOLD`	0.72–0.78	It is necessary to distinguish similar heating areas and campaign types. A value that is too low will introduce irrelevant results, while a value that is too high will miss accurately matched content
`INCREMENTAL_INDEX_ENABLE`	Enabled	Heating marketing content updates are concentrated and added on demand. Incremental indexing avoids the resource consumption of full retraining
`DEDUPLICATE_CHUNKS`	Enabled	Avoid duplicate segments caused by the `chunk_overlap` parameter, reducing index redundancy

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

Three Common Configuration Mistakes

Phenomenon: After uploading a single heating marketing document, the number of parsed segments is shown as 13, while the actual document is divided into 8 segments, with duplicate segments appearing. Cause: The DEDUPLICATE_CHUNKS switch is not enabled, and the chunk_overlap parameter is set too large, causing text in overlapping intervals to be repeatedly counted in different segments.
Phenomenon: After upgrading from version 4.9.0 to 4.9.3, existing heating marketing content cannot be recalled. Cause: The vector model encoding logic of the new version differs from the old version. If the vector index is not regenerated, the original index cannot match new query embeddings.
Phenomenon: A timeout error is triggered when batch updating heating marketing content. Cause: INCREMENTAL_INDEX_ENABLE is not enabled. When performing full retraining, the PARSE_FILE_TIMEOUT_SECONDS parameter is not adjusted, causing parsing timeout for large individual files.

How to Verify Correct Configuration

Upload a typical heating marketing campaign copy, view the parsed segment list, check the match between the number of segments and the document structure, and adjust chunk_size and chunk_overlap until there is no obvious semantic break in the segments.
Initiate a simulated user query, such as "XX street heating payment discount", check the similarity scores of recall results, and adjust SIMILARITY_THRESHOLD until the results accurately match the query scenario.
Upload an updated marketing campaign copy, check the index task status, confirm that only new segments are processed and no full retraining is triggered, verifying that the incremental indexing configuration is effective.
Batch upload multiple identical types of marketing documents, check the total number of segments after indexing is completed, confirm that there are no duplicate segments, verifying that the DEDUPLICATE_CHUNKS configuration is effective.

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.