Defines the `EmbedService` trait with a single method to embed text into vectors. Provides a factory function `create_embed_service` that instantiates concrete implementations (Bedrock, OpenAI-compatible, Noop) from configuration. This pattern is self-contained, decouples embedding logic from the rest of the system, and can be extracted into a new project to support multiple embedding providers with minimal adaptation.

Analogous to the embed service, defines the `LlmService` trait for LLM calls (chat completion, embedding, etc.). The factory creates instances from config, supporting AWS Bedrock, OpenAI-compatible APIs, and a Noop provider. This modular design allows any LLM-dependent module to remain agnostic of the underlying provider, and the pattern is directly reusable in other Rust applications requiring abstracted LLM access.

Implements a cross-encoder-based reranker that takes a list of search results, scores them with a cross-encoder model, and reapplies a configurable threshold. The module is self-contained, with its own configuration and clear input/output (a list of scored `Memory` records). It demonstrates non-obvious skill in handling the reranking stage with fallback to original scores and a configurable cutoff, making it a reusable component for any retrieval pipeline that needs precision improvement.