Coverage for src/lilbee/core/llm

1"""Extract the first JSON value from an LLM reply.

3Models wrap JSON in prose, fences, and trailing commentary. A greedy

4``\\{.*\\}`` span runs to the LAST brace in the reply, so any trailing text

5containing one drops the whole value; a hand-rolled brace counter miscounts

6braces inside string literals. ``raw_decode`` scanned from each opener lets the

7stdlib own the parsing state, which is the only thing that gets both right.

8"""

10from __future__ import annotations

12import json

13from typing import TypeVar

15_DECODER = json.JSONDecoder()

18def json_reply_format() -> dict[str, str]:

19 """Provider option asking for a bare JSON value rather than JSON in prose.

21 llama.cpp's server and OpenAI-compatible remotes both honour it. Returns

22 a fresh dict per call because it is handed to a third-party SDK that is free

23 to mutate the request it is given. The scans below stay the fallback: a

24 provider can ignore the request and a model can comply imperfectly, and

25 neither should cost the caller its answer.

26 """

27 return {"type": "json_object"}

30_JsonT = TypeVar("_JsonT", dict, list)

33def _first_json_value(text: str, opener: str, expected: type[_JsonT]) -> _JsonT | None:

34 """The first *expected*-typed JSON value starting at an *opener*, or None."""

35 start = text.find(opener)

36 while start >= 0:

37 try:

38 parsed, _ = _DECODER.raw_decode(text, start)

39 except json.JSONDecodeError:

40 start = text.find(opener, start + 1)

41 continue

42 return parsed if isinstance(parsed, expected) else None

43 return None

46def first_json_object(text: str) -> dict | None:

47 """The first JSON object in *text*, or None."""

48 return _first_json_value(text, "{", dict)

51def first_json_array(text: str) -> list | None:

52 """The first JSON array in *text*, or None."""

53 return _first_json_value(text, "[", list)

Coverage for src/lilbee/core/llm_json.py: 100%