fix: guard malformed JSON in JSONConverter content_key path

[chuenchen309]

What

JSONConverter._extract_text_and_metadata guards the jq_schema path (it logs a warning and skips the file on failure), but the non-jq branch — used when the converter is configured with content_key and no jq_schema — calls json.loads(file_content) with no guard:

else:
    # We just load the whole file as JSON if the user didn't provide a jq filter.
    objects = [json.loads(file_content)]

A single malformed .json source therefore raises an unhandled json.JSONDecodeError and aborts the entire run(), even when other sources in the batch are valid. This is inconsistent with the jq_schema path and with the existing handling of decode errors (UnicodeError) a few lines above, both of which skip the offending file and continue.

Fix

Wrap the non-jq json.loads in the same warn-and-skip guard already used by the jq branch, catching the specific json.JSONDecodeError.

Test

Added test_run_with_malformed_json_and_content_key, which mirrors the existing test_run_with_non_json_file but exercises the content_key-only path. It fails before the change (uncaught JSONDecodeError) and passes after, asserting the file is skipped with a single warning and run() returns {"documents": []}.

Verified locally: the new test and the other non-jq tests in test/components/converters/test_json.py pass; the jq-dependent tests require the optional jq extra (not installed locally) and are unaffected by this change. ruff check and ruff format --check are clean. Added a release note.

---

Authored with AI assistance (Claude Code); the change and tests were reviewed and verified locally by me.

[vercel]

@chuenchen309 is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
haystack/components/converters
json.py
Project Total

This report was generated by python-coverage-comment-action

[davidsbatista]

Hi @chuenchen309, thanks for the opening the PR.

I also noticed you opened 7 PRs in the last few hours. Was this due to you using Haystack or by running AI tools on the code?

[chuenchen309]

Hi @davidsbatista — happy to be straight about it. I'm a light user of Haystack, not a heavy day-to-day one, but it's a project I'm genuinely interested in and wanted to contribute back to. For that reason, most of these came from using an AI coding assistant (Claude Code) to do focused code review across parts of the codebase, rather than from bugs I stumbled onto purely through my own usage — that's why several landed in a short window. Each PR isn't just an AI patch pasted in though: for every one I had it write a failing test that reproduces the actual defect, verified the fix makes the test pass, ran the relevant suite plus ruff/fmt locally, and read the diff myself before opening it, so I stand behind the correctness of each fix.

I take the point on pace — 7 PRs in a few hours is a lot to land on anyone's plate. I'll hold off on opening any more for now so these four don't get buried; happy to revisit later if useful, and no worries at all if some/all of these aren't a good fit for the project right now.

[davidsbatista]

Hi @davidsbatista — happy to be straight about it. I'm a light user of Haystack, not a heavy day-to-day one, but it's a project I'm genuinely interested in and wanted to contribute back to. For that reason, most of these came from using an AI coding assistant (Claude Code) to do focused code review across parts of the codebase, rather than from bugs I stumbled onto purely through my own usage — that's why several landed in a short window. Each PR isn't just an AI patch pasted in though: for every one I had it write a failing test that reproduces the actual defect, verified the fix makes the test pass, ran the relevant suite plus ruff/fmt locally, and read the diff myself before opening it, so I stand behind the correctness of each fix.

I take the point on pace — 7 PRs in a few hours is a lot to land on anyone's plate. I'll hold off on opening any more for now so these four don't get buried; happy to revisit later if useful, and no worries at all if some/all of these aren't a good fit for the project right now.

Thanks a lot for the clarification, and happy to know about your interest and use of Haystack, also that not all PRs were automated and that some human reasoning was involved. We will review the PRs as time permits.

[chuenchen309]

Thanks @davidsbatista, really appreciate it — no rush on any of these. 🙏

[davidsbatista]

Looks good! Thanks for the contribution.