zylon-ai/private-gpt · error · ValueError

Query not found: {sql_query}

Error message

Query not found: {sql_query}

What it means

Raised by the execute_sql_query helper that _process_sql_queries injects into the sandbox: every SQL query found in the generated code is pre-executed and its result written to a CSV keyed by the query string in _datasets_map. At runtime the generated code calls execute_sql_query(sql) and a lookup miss (exact string mismatch) raises this ValueError inside the sandbox. It almost always means the query string used at runtime differs from the one extracted at pre-processing time.

Source

Thrown at private_gpt/components/tabular/pandasai_sandbox.py:343

            except Exception as e:
                custom_exception = e.__class__(clean_exception_text(str(e)))
                exceptions.append(custom_exception)
                logger.error("Failed to execute SQL query: %s", custom_exception)

        if not datasets_map and exceptions:
            return "", exceptions

        datasets_code = textwrap.dedent(
            f"""
            import os
            import pandas as pd
            _datasets_map = {datasets_map!r}
            _temp_dir = {temp_dir!r}
            def execute_sql_query(sql_query):
                filename = _datasets_map.get(sql_query)
                if filename:
                    return pd.read_csv(os.path.join(_temp_dir, filename))
                raise ValueError(f'Query not found: {{sql_query}}')
        """
        ).strip()

        return datasets_code, []

    def _prepare_code_for_execution(self, code: str) -> str:
        temp_dir = f"/tmp/{self._user_id}"

        # Redirect any hardcoded .png paths into the sandbox temp dir
        code = re.sub(
            r"""(['"])([^'"]*\.png)\1""",
            lambda m: (
                f"{m.group(1)}{temp_dir}/{os.path.basename(m.group(2))}{m.group(1)}"
            ),
            code,
        )

        # Replace explicit color lists with CUSTOM_COLORS

View on GitHub (pinned to 4a030776a3)

Solutions

1. Log both the map keys (_datasets_map) and the failing lookup string to see the exact mismatch.
2. Normalize map keys and lookups (strip/collapse whitespace) before comparison.
3. Prefer literal SQL strings in generated code — prompt the model not to construct SQL dynamically.
4. As a mitigation, fall back to executing the query directly when the lookup misses instead of raising.

Example fix

# before
datasets_code = textwrap.dedent(
    f"""
    def execute_sql_query(sql_query):
        filename = _datasets_map.get(sql_query)
        if filename:
            return pd.read_csv(os.path.join(_temp_dir, filename))
        raise ValueError(f'Query not found: {{sql_query}}')
    """
)

# after (normalized lookup)
_datasets_map_norm = {{k.strip(): v for k, v in _datasets_map.items()}}
def execute_sql_query(sql_query):
    filename = _datasets_map_norm.get(sql_query.strip())
    if filename:
        return pd.read_csv(os.path.join(_temp_dir, filename))
    raise ValueError(f'Query not found: {sql_query}')

# validate before execution: every SQL literal in generated code must be extractable
import re
SQL_RE = re.compile(r'(['"])(SELECT .*?FROM .*?)\1', re.IGNORECASE | re.DOTALL)

def sql_literals_are_static(code: str) -> bool:
    return not re.search(r'execute_sql_query\s*\(\s*f?["\'].*\+', code)

When it happens

Trigger: The generated code builds the SQL string dynamically (f-string/concatenation) so the runtime string differs from the literal seen by _extract_sql_queries_from_code; string escaping differences between extraction and execution; whitespace/quoting differences between the extracted query and the map key.

Common situations: LLM writes parameterized or templated SQL; the extraction regex captures a slightly different span than what the code later passes; duplicate near-identical queries with subtle whitespace differences.

Related errors

• Failed to execute some SQL queries: {', '.join(str(e) for e
• Invalid CALL statement format
• Code execution failed: {message}
• Path '{canonical_path}' does not match any session mount.
• Failed to parse JSON: {e!s}

AI-assisted analysis of zylon-ai/private-gpt@4a030776a3 (2026-08-15). Data as JSON: /api/errors/7cc860f6550e4cb8. Report an issue: GitHub.