✨ new implement : support Rerank models in model management and agent configurations#2700
✨ new implement : support Rerank models in model management and agent configurations#2700
Conversation
Codecov Report❌ Patch coverage is 📢 Thoughts on this report? Let us know! |
| kb_search_results.extend(kb_search) | ||
|
|
||
| # Apply reranking if enabled | ||
| if self.rerank and self.rerank_model and kb_search_results: |
There was a problem hiding this comment.
知识库检索支持topk参数，即返回最接近结果的前k个检索结果。
之前检索是 每个知识库都返回k个知识库结果。返回的是k*n（n是知识库数量）
现在有了reranker之后，可以将所有检索结果综合排序，返回k个结果
| @@ -93,6 +104,9 @@ def __init__( | |||
| self.top_k = top_k | |||
There was a problem hiding this comment.
对于 没有开启 rerank的检索工具，topk就返回k个检索结果。
开启了reranker，那么检索结果 应是 k* 系数个检索结果，然后从总的检索结果，综合排序 选择最优。
这个系数 我去确认下。
There was a problem hiding this comment.
可以先按照10倍，即开启rerank，先检索 10*topK个检索结果
There was a problem hiding this comment.
系数设定为10，可在nexent/sdk/nexent/core/utils/constants.py 中通过修改RERANK_OVERSEARCH_MULTIPLIER更改
| default="", | ||
| ), | ||
| observer: MessageObserver = Field( | ||
| description="Message observer", default=None, exclude=True), |
There was a problem hiding this comment.
对于每个工具来说，还提供了工具测试功能。所以，还需要保证加了rerank功能后，测试工具功能也可以正常运行
There was a problem hiding this comment.
3个使用到重排模型的工具，重排功能开启，工具测试均通过
wadecrack
force-pushed
the
xq/develop_import_reranker_model
branch
from
1c39c1e to
1ce40ff
Compare
|
wadecrack
force-pushed
the
xq/develop_import_reranker_model
branch
from
abb7061 to
ce77cc3
Compare
2 participants
Added Rerank model support to the Model Management page (supports batch adding).
New Rerank toggle available for search tools (datamate, knowledge_base, dify) in Agent Space.
Enabling Rerank allows for secondary processing of retrieved chunks, leading to higher-quality AI answers in knowledge-based scenarios.