[Feature]: Ideas for Improvement for docs/dynamic-tool-discovery.md

Ideas for Dynamic Tool Discovery Improvements:

- Hybrid Search with Vector Search + BM25 yields better results than just vector search. Suggested DB with HybridSearch support: [LanceDB](https://github.com/lancedb/lancedb), [VectorChord](https://docs.vectorchord.ai/vectorchord/getting-started/vectorchord-suite.html#how-to-use-the-vectorchord-suite)
- Saving & re-using vectors for the processed MCP Tools in the VectorDB
- Semantic Pre-filtering of the enums in the tool, for example, when a timezone tool has a long enumeration list of all available timezones around the globe, it makes sense to pre-filter the enum list and extract only a few relevant timezone examples, saving LLM context and helping LLM focus. Proposed [algorithm for enum reduction](https://github.com/stanford-mast/a1/blob/main/src/a1/extra_strategies.py)
- I'd like to be able to enable intelligent_tool_finder to be **_always_** invoked **_before the LLM call_**, using the original unmodified user prompt as a search query with the semantic search
- I'd like to enable the Threshold Semantic Similarity score instead of TopK/TopN, so my relevant list of tools would not be cut off
- Long user queries makes sence to semantically chunk into smaller portions of text and search tools for each chunk. Suggested chunking framework: [Chonkie](https://docs.chonkie.ai/oss/chunkers/semantic-chunker)
- MCP Server re-scan & re-index based on events or schedule on background, and granular embedding re-generation only for the updated tools. Preserve the previous versions in the DB in case of flapping

The goal is to further improve accuracy and speed with "cheap" techniques, caching, removing unnecessary LLM steps, and some parameter tuning.