Use JSON Patch instead of entire objects for retries

[NickSherrow]

Hey everyone,

The current retry logic requires the LLM to regenerate the entire object every time the received object fails validation. I think this might not be the best approach for fixing objects, and here’s why:

• For complex schemas, this can be problematic: while the LLM is fixing one field, it might break another. As a result, a valid object may never be produced, even after 20–30 attempts.
• The object size is constrained by the model’s output token window.
• This approach scales poorly in terms of token usage.

Why not use something like [RFC 6902 JSON Patch](https://tools.ietf.org/html/rfc6902), as [trustcall](https://github.com/hinthornw/trustcall) does?

An alternative flow could look like this:

• Define a Patch class with op, path, and value, generated dynamically from the schema.
• If validation fails, ask the LLM to generate a set of patches. If that fails, retry.
• Apply patches, re-validate, and repeat as needed.

Has this approach already been discussed? Curious what the community thinks about this in general.