run-llama/llama_index · error · OutputParserException

Got invalid JSON object. Error: {e_json} {e_yaml}. Got JSON

Error message

Got invalid JSON object. Error: {e_json} {e_yaml}. Got JSON string: {json_string}

What it means

SelectionOutputParser.parse (used for LLM selection/routing output) first tries json.loads on the LLM string, then falls back to yaml.safe_load (which tolerates trailing commas). Only when BOTH parsers fail does it raise OutputParserException embedding both error messages and the offending string — i.e. the model's output was neither valid JSON nor YAML.

Source

Thrown at llama-index-core/llama_index/core/output_parsers/selection.py:86

            output_json.append(json_dict)

        return output_json

    def parse(self, output: str) -> Any:
        json_string = _marshal_llm_to_json(output)
        try:
            json_obj = json.loads(json_string)
        except json.JSONDecodeError as e_json:
            try:
                import yaml

                # NOTE: parsing again with pyyaml
                #       pyyaml is less strict, and allows for trailing commas
                #       right now we rely on this since guidance program generates
                #       trailing commas
                json_obj = yaml.safe_load(json_string)
            except yaml.YAMLError as e_yaml:
                raise OutputParserException(
                    f"Got invalid JSON object. Error: {e_json} {e_yaml}. "
                    f"Got JSON string: {json_string}"
                )
            except NameError as exc:
                raise ImportError("Please pip install PyYAML.") from exc

        if isinstance(json_obj, dict):
            json_obj = [json_obj]

        if not isinstance(json_obj, list):
            raise ValueError(f"Failed to convert output to JSON: {output!r}")

        json_output = self._format_output(json_obj)
        answers = [Answer.from_dict(json_dict) for json_dict in json_output]
        return StructuredOutput(raw_output=output, parsed_output=answers)

    def format(self, prompt_template: str) -> str:
        return prompt_template + "\n\n" + _escape_curly_braces(FORMAT_STR)

View on GitHub (pinned to afd0fef371)

Solutions

1. Catch OutputParserException and retry the query or re-ask with a stricter instruction
2. Raise max_tokens/num_output so the selection payload is never truncated
3. Use a stronger model or structured-output-capable model for routing/selection
4. Sanitize model output before parsing (strip markdown fences, smart quotes) via a pre-processing wrapper

Example fix

# before
response = router_query_engine.query("What is 2+2?")  # OutputParserException: invalid JSON/YAML

# after
from llama_index.core.output_parsers import OutputParserException
try:
    response = router_query_engine.query("What is 2+2?")
except OutputParserException:
    response = router_query_engine.query("Answer using the required JSON schema exactly.")

import json
def is_parseable_selection_output(s: str) -> bool:
    try:
        json.loads(s)
        return True
    except json.JSONDecodeError:
        try:
            import yaml
            obj = yaml.safe_load(s)
            return isinstance(obj, (dict, list))
        except Exception:
            return False

from llama_index.core.output_parsers import OutputParserException
for attempt in range(2):
    try:
        result = selector.select(prompt)
        break
    except OutputParserException as e:
        if attempt == 1:
            raise
        prompt = prompt + "\nRespond ONLY with valid JSON (no prose, no trailing commas beyond YAML tolerance)."

When it happens

Trigger: An LLM (via RouterQueryEngine/selection prompting) returning output with unterminated strings, smart quotes, unescaped newlines in values, markdown fences mixed with truncation, or truncated output exceeding max_tokens; also any string that is invalid in both grammars (e.g. a bare sentence with a colon, YAML may still parse — truly invalid: unbalanced brackets).

Common situations: Weaker/open models producing malformed selection JSON; max_tokens too small so choices array is cut off; prompts altered so the model answers in prose; tool descriptions containing braces that confuse the model's formatting.

Understand the failure class

• Parsing and encoding errors: unexpected token, malformed input — why parsers reject input and how to find the real culprit.

Related errors

• Failed to convert output to JSON: {output!r}
• Please pip install PyYAML.
• Failed to parse pydantic object from guidance program. Proba
• Got empty streaming response
• Expected ActionReasoningStep, got {reasoning_step}

AI-assisted analysis of run-llama/llama_index@afd0fef371 (2026-08-15). Data as JSON: /api/errors/b3251f723ac9d587. Report an issue: GitHub.