refactor: improve time complexity and reduce memory allocations

[chabinhwang]

Summary

• DefaultToolCallingManager: Pre-compute HashSet of existing tool names to reduce resolveToolDefinitions lookup from O(N×M) to O(N+M)
• ModelOptionsUtils: Change interfaceNames from List to Set for O(1) contains() instead of O(N), use Character.toUpperCase() to avoid unnecessary String allocation in toGetName()
• Vector stores (16 files): Replace documents.indexOf(document) O(N) lookup inside loops with index-based access O(1), eliminating O(N²) overall complexity in doAdd methods

Details

DefaultToolCallingManager (resolveToolDefinitions)

Previously, for each tool name in chatOptions.getToolNames(), the code streamed through all toolCallbacks with .anyMatch() — resulting in O(N×M) complexity. Now a HashSet<String> of existing tool callback names is pre-computed once, reducing the check to O(1) per iteration.

ModelOptionsUtils (mergeBeans / toGetName)

• interfaceNames was collected as a List<String> but used with .contains() which is O(N) per call. Changed to Set<String> for O(1) lookups.
• toGetName() used name.substring(0, 1).toUpperCase() which creates an intermediate String object. Changed to Character.toUpperCase(name.charAt(0)) to avoid unnecessary allocation.

Vector Stores (16 files)

All vector store doAdd methods used documents.indexOf(document) inside a for-each loop to correlate documents with their embeddings. Since indexOf is O(N), this made the overall loop O(N²). Replaced with indexed for-loops using embeddings.get(i) for O(1) access.

Affected stores: Cassandra, Chroma, Couchbase, GemFire, MariaDB, Milvus, MongoDB Atlas, Neo4j, OpenSearch, Oracle, Pinecone, Qdrant, Redis, Typesense, Azure, CosmosDB