[特性]: knowledge-search 将普通文档标签过滤下推到向量索引，避免先展开 knowledge_ids

作者: songwenyong创建于 2026年9月17日更新于 2026年9月17日

The current knowledge index construction does not write to the TagID. The document tags need to be synchronized during the storage, reanalysis, and related index rebuilding paths.
Knowledge documents already support multiple tags, and it is not possible to simply copy one tag to a single-value tag_id and lose the other tags. It is possible to use the multi-value tag field supported by the backend (such as tag_ids) or equivalent index representations to preserve the existing OR semantics of "matching any tag"; the existing single-value tag_id for FAQ is also supported.
When a document tag is added, modified, removed, or a tag is deleted, the search index metadata should be synchronized, and clear failure retries and consistency handling should be specified; it should not be necessary to recalculate the text vector for each tag modification.
Historical indexes need to be backfilled or migrated; the old path should be retained as a compatibility fallback to avoid sudden unavailability of results for existing knowledge bases. Backends that do not support corresponding metadata filtering can also retain the fallback.

内容来源: Tencent/WeKnora