perf: Defer query-time imports in conversation_base

[KRRT7]

Stack: depends on #235. Merge #235 first, then this PR.

---

conversation_base.py eagerly imported four query-time modules at module level:

• answers (20 ms cumulative — pulls in search + query)
• search_query_schema (10 ms)
• searchlang (3 ms)
• answer_response_schema (1 ms)

These are only used in the query() method. Moved the imports inside query() and used TYPE_CHECKING + from __future__ import annotations for the type hints.

Benchmark

Azure Standard_D2s_v5 — 2 vCPU, 8 GiB RAM, Python 3.13

Import Time (hyperfine, warmup 5, min-runs 30)

Benchmark	Before (#235)	After	Speedup
import typeagent	734 ms ± 13 ms	683 ms ± 15 ms	1.07x

Cumulative since baseline (main):

Benchmark	main	This PR	Speedup
import typeagent	791 ms ± 11 ms	683 ms ± 15 ms	1.16x

---

Upstream opportunity: pydantic_ai transitive imports

pydantic_ai contributes ~161 ms cumulative to import typeagent through its transitive dependency chain:

Module	Cumulative (ms)	Notes
logfire	43	Observability SDK, imported eagerly
griffe	34	Code introspection for tool schemas
pydantic_ai.messages	22	Pydantic model definitions
genai_prices	9	Pricing data snapshot

This is the single largest remaining import cost and would need to be addressed upstream in pydantic-ai.

---

Generated by codeflash optimization agent

[gvanrossum]

I don't follow precisely what games you are playing with from __future__ import annotations, but I don't want it in my code -- while it's not yet deprecated, eventually it will be, and I don't want to add a future burden. If that means the whole optimization can't happen, so be it -- as I wrote before, these optimizations aren't all that valuable given the typical LLM roundtrip.

[KRRT7]

Good point — I can drop from __future__ import annotations and use string literals for the four annotations that reference the deferred modules instead:

_query_translator: "typechat.TypeChatJsonTranslator[search_query_schema.SearchQuery] | None" = None
_answer_translator: "typechat.TypeChatJsonTranslator[answer_response_schema.AnswerResponse] | None" = None

and in the query() signature:

async def query(
    self,
    question: str,
    search_options: "searchlang.LanguageSearchOptions | None" = None,
    answer_options: "answers.AnswerContextOptions | None" = None,
) -> str:

The TYPE_CHECKING block + runtime import inside query() stay the same — only the future import goes away. Type checkers still resolve the string annotations against the TYPE_CHECKING imports. I'll push that update.

On the value question — I was honestly hesitant to open import-time optimizations at all, which is why I asked first on #229 before going further. That said, if you'd rather keep the code straightforward and not defer these, I'm happy to close this one.