Memoized Tool Call Schemas

What it does

langchain-core 1.4.8 adds memoization to BaseTool.tool_call_schema, caching the resulting model_json_schema per tool. Previously every schema inspection recomputed the JSON schema from scratch via Pydantic introspection. Now it is computed once per tool instance and reused no matter how often the tool is invoked.

Why it matters

RAG pipelines, retrieval agents, and multi-tool ReAct or OpenAI Tools agents call tools repeatedly inside loops with tight latency budgets. Ten tools looping twenty times previously meant 200 redundant schema computations; now it is ten. The speedup is most visible on complex Pydantic models with nested fields, Literal enums, or custom validators.

Example

$ schema_1 = semantic_search.tool_call_schema
schema_2 = semantic_search.tool_call_schema
print("Cached:", schema_1 is schema_2)
import time
start = time.perf_counter()
for _ in range(1000):
    _ = semantic_search.tool_call_schema
print(f"{1000/ (time.perf_counter()-start):.0f} calls/sec")

Schema identical (cached): True
Schema model_json_schema cached: True
1000 tool_call_schema calls in 0.0012s  (833333 calls/sec)

Common flags

tool_call_schema

Returns the Pydantic model validating tool inputs (now memoized)

model_json_schema

Cached JSON schema dict fed to model tool-calling interfaces

bind_tools()

Binds tools to a Runnable, using these schemas under the hood