fix(nodes): tolerate doubled-brace JSON output from models like DeepSeek

[mjmirza]

Summary

GenerateAnswerNode fails with OutputParserException: Invalid json output when the LLM returns its JSON wrapped in doubled braces, e.g. {{"content": "..."}}. This happens reliably with DeepSeek (deepseek/deepseek-chat) and intermittently with other less strict models.

Root cause

The schema-less format_instructions show the expected shape as:

{{"content": "your analysis here"}}

Those doubled braces are LangChain template escaping for a single literal { }. GPT-4o-class models follow the instruction and emit single braces, but some models (notably DeepSeek) copy the doubled braces verbatim into their answer, which is not valid JSON, so JsonOutputParser raises.

Fix

Add TolerantJsonOutputParser, a small JsonOutputParser subclass that, only on the parse-failure path, retries once with a single layer of wrapping braces removed. Behaviour is unchanged for any model already returning valid JSON (the happy path goes straight through the parent parser). It is used in the schema-less branch of GenerateAnswerNode.

• scrapegraphai/utils/output_parser.py — new TolerantJsonOutputParser + _strip_doubled_braces helper
• scrapegraphai/nodes/generate_answer_node.py — use it in the non-schema branch
• tests/utils/output_parser_test.py — 7 unit tests (clean JSON unchanged, doubled-brace recovery, whitespace, irrecoverable output still raises)

Reproduction (before this change)

from scrapegraphai.graphs import SmartScraperGraph
cfg = {"llm": {"api_key": "<deepseek-key>", "model": "deepseek/deepseek-chat"}, "headless": True}
SmartScraperGraph(prompt="Extract the heading", source="https://example.com", config=cfg).run()
# -> OutputParserException: Invalid json output: {{"content": "..."}}

After: returns the parsed dict as expected.

Testing

python -m pytest tests/utils/output_parser_test.py
# 7 passed

Suggested labels: bug, size:S

[github-actions]

🎉 This PR is included in version 2.2.0-beta.4 🎉

The release is available on:

• v2.2.0-beta.4
• GitHub release

Your semantic-release bot 📦🚀