Describe your use case
EN. Retrieval (src/contextseek/retrieval/orchestrator.py +
components.py) does hybrid keyword+vector recall with optional LLM reranking.
Introduce a pluggable reranker interface and add a non-LLM option — a
cross-encoder (e.g. bge-reranker via sentence-transformers) or a hosted rerank
API — selectable by config. This gives users a cheaper/faster alternative to LLM
reranking.
Tasks: define a Reranker protocol; refactor the existing LLM reranker to
implement it; add the new reranker; make the choice configurable in
config/settings.py; benchmark recall/latency in a short note; add tests with a
stub reranker.
Acceptance: reranker is swappable via config; the LLM path is unchanged in
behavior; new path has tests.
ZH. 检索（src/contextseek/retrieval/orchestrator.py + components.py）做的
是关键词+向量混合召回，外加可选的 LLM rerank。请引入一个可插拔的 reranker 接
口，并加一个非 LLM 选项——cross-encoder（如经 sentence-transformers 的
bge-reranker）或托管 rerank API——可经配置选择。给用户一个比 LLM rerank 更便宜/更
快的替代。
任务： 定义 Reranker 协议；把现有 LLM reranker 重构为实现该协议；加新
reranker；在 config/settings.py 让选择可配置；用一段简短说明给出召回/延迟对比；用
桩 reranker 写测试。
验收： reranker 可经配置替换；LLM 路径行为不变；新路径有测试。
Describe the solution you'd like
Pluggable reranker in the retrieval orchestrator
Describe alternatives you've considered
No response
Additional context
No response
Describe your use case
EN. Retrieval (
src/contextseek/retrieval/orchestrator.py+components.py) does hybrid keyword+vector recall with optional LLM reranking.Introduce a pluggable reranker interface and add a non-LLM option — a
cross-encoder (e.g.
bge-rerankervia sentence-transformers) or a hosted rerankAPI — selectable by config. This gives users a cheaper/faster alternative to LLM
reranking.
Tasks: define a
Rerankerprotocol; refactor the existing LLM reranker toimplement it; add the new reranker; make the choice configurable in
config/settings.py; benchmark recall/latency in a short note; add tests with astub reranker.
Acceptance: reranker is swappable via config; the LLM path is unchanged in
behavior; new path has tests.
ZH. 检索（
src/contextseek/retrieval/orchestrator.py+components.py）做的是关键词+向量混合召回，外加可选的 LLM rerank。请引入一个可插拔的 reranker 接
口，并加一个非 LLM 选项——cross-encoder（如经 sentence-transformers 的
bge-reranker）或托管 rerank API——可经配置选择。给用户一个比 LLM rerank 更便宜/更快的替代。
任务： 定义
Reranker协议；把现有 LLM reranker 重构为实现该协议；加新reranker；在
config/settings.py让选择可配置；用一段简短说明给出召回/延迟对比；用桩 reranker 写测试。
验收： reranker 可经配置替换；LLM 路径行为不变；新路径有测试。
Describe the solution you'd like
Pluggable reranker in the retrieval orchestrator
Describe alternatives you've considered
No response
Additional context
No response