Improve LLM error handling for structured output

[gfargo]

Describe the request

Sometimes the LLM will respond with a malformed object when using structure output with LangChain. To compensate for this, we should use Zod to validate the response.

Additionally, we can add a maxRetry config value that will automatically attempt to retry the request until response matches the expected shape.

[gfargo]

Note solution used in #292 to ensure only a JSON object was returned, sans prefixed intro LLM text

[gfargo]

• https://js.langchain.com/docs/concepts/output_parsers
• https://python.langchain.com/docs/how_to/structured_output/
• https://python.langchain.com/docs/how_to/output_parser_json/
  • https://v03.api.js.langchain.com/classes/langchain.output_parsers.OutputFixingParser.html