|
1
|
"""Robust JSON extraction from LLM responses.""" |
|
2
|
|
|
3
|
import json |
|
4
|
import re |
|
5
|
from typing import Optional, Union |
|
6
|
|
|
7
|
|
|
8
|
def parse_json_from_response(text: str) -> Optional[Union[list, dict]]: |
|
9
|
""" |
|
10
|
Extract JSON from an LLM response, handling markdown fences, |
|
11
|
explanatory text, and minor formatting issues. |
|
12
|
|
|
13
|
Strategies tried in order: |
|
14
|
1. Direct parse |
|
15
|
2. Strip markdown fences and parse |
|
16
|
3. Find [...] or {...} substring and parse |
|
17
|
4. Return None |
|
18
|
""" |
|
19
|
if not text or not text.strip(): |
|
20
|
return None |
|
21
|
|
|
22
|
cleaned = text.strip() |
|
23
|
|
|
24
|
# Strategy 1: direct parse |
|
25
|
try: |
|
26
|
return json.loads(cleaned) |
|
27
|
except json.JSONDecodeError: |
|
28
|
pass |
|
29
|
|
|
30
|
# Strategy 2: strip markdown fences |
|
31
|
fence_pattern = re.compile(r"```(?:json)?\s*\n?(.*?)\n?\s*```", re.DOTALL) |
|
32
|
match = fence_pattern.search(cleaned) |
|
33
|
if match: |
|
34
|
try: |
|
35
|
return json.loads(match.group(1).strip()) |
|
36
|
except json.JSONDecodeError: |
|
37
|
pass |
|
38
|
|
|
39
|
# Strategy 3: find JSON array or object |
|
40
|
# Try array first (often the outermost structure for lists) |
|
41
|
for opener, closer in [("[", "]"), ("{", "}")]: |
|
42
|
start = cleaned.find(opener) |
|
43
|
if start < 0: |
|
44
|
continue |
|
45
|
# Find matching closer (handle nesting) |
|
46
|
depth = 0 |
|
47
|
for i in range(start, len(cleaned)): |
|
48
|
if cleaned[i] == opener: |
|
49
|
depth += 1 |
|
50
|
elif cleaned[i] == closer: |
|
51
|
depth -= 1 |
|
52
|
if depth == 0: |
|
53
|
try: |
|
54
|
return json.loads(cleaned[start : i + 1]) |
|
55
|
except json.JSONDecodeError: |
|
56
|
break |
|
57
|
|
|
58
|
return None |
|
59
|
|