PR #36748: fix(core): parse_partial_json escapes raw \r and \t inside strings (#36747)

• Repository: langchain-ai/langchain
• Author: MukundaKatta
• State: closed | merged: False
• Link: https://github.com/langchain-ai/langchain/pull/36748

Description (problem / solution / changelog)

Why

Closes #36747.

parse_partial_json already escapes raw \n inside JSON string values (required by RFC 8259 — unescaped control chars are illegal in JSON strings). But \r and \t were not handled, so model outputs containing those raw control chars would crash the parser with json.JSONDecodeError: Invalid control character.

Repro from the issue:

>>> parse_partial_json('{"key": "col1\tcol2"}')
# raises json.JSONDecodeError

What

Two more elif branches in the inner character loop, mirroring the existing \n handler:

elif char == "\r" and not escaped:
    new_char = "\\r"
elif char == "\t" and not escaped:
    new_char = "\\t"

The sibling helper _replace_new_line already handles all three in the non-streaming path (line 25-27), so this just brings the streaming parser up to the same behavior.

Tests

4 new parametrized cases in TEST_CASES_PARTIAL:

('{"key": "line1\nline2"}', '{"key": "line1\\nline2"}'),   # pre-existing behavior, now explicit
('{"key": "line1\rline2"}', '{"key": "line1\\rline2"}'),   # #36747 regression guard
('{"key": "col1\tcol2"}', '{"key": "col1\\tcol2"}'),       # #36747 regression guard
('{"k": "a\nb\rc\td"}', '{"k": "a\\nb\\rc\\td"}'),         # mixed

$ pytest libs/core/tests/unit_tests/output_parsers/test_json.py -k parse_partial -v
13 passed

Scope

Minimal behavior change. No public API change. Doesn't affect any other JSON parsing path — only the character-by-character recovery loop inside parse_partial_json.

Related

• Closes #36747
• Aligns with _replace_new_line helper (same file) which already handles all three control chars

Changed files

• libs/core/langchain_core/utils/json.py (modified, +5/-3)
• libs/core/tests/unit_tests/output_parsers/test_json.py (modified, +8/-0)

PR #36763: fix(core): escape all JSON control characters in parse_partial_json

• Repository: langchain-ai/langchain
• Author: Dheeraj-Bhaskaruni
• State: closed | merged: False
• Link: https://github.com/langchain-ai/langchain/pull/36763

Description (problem / solution / changelog)

Description

Fixes #36747

parse_partial_json only escaped raw \n inside JSON strings but not other control characters that are illegal per RFC 8259 (U+0000 through U+001F). LLMs can emit raw \r, \t, \b, \f, and other control characters in their output, causing json.JSONDecodeError.

Changes

• langchain_core/utils/json.py: Added escaping for \r, \t, \b, \f, and a catch-all for remaining control characters using \uXXXX encoding
• tests/unit_tests/output_parsers/test_json.py: Added 6 test cases covering \r, \t, \b, \f, null char, and mixed control characters

Before

parse_partial_json('{"key": "line1\rline2"}')
# raises json.JSONDecodeError: Invalid control character

After

parse_partial_json('{"key": "line1\rline2"}')
# {'key': 'line1\rline2'}

Changed files

• libs/core/langchain_core/utils/json.py (modified, +13/-3)
• libs/core/tests/unit_tests/output_parsers/test_json.py (modified, +6/-0)