xebia-functional / xef

Building applications with LLMs through composability, in Kotlin, Scala, ...
https://xef.ai
Apache License 2.0
174 stars 15 forks source link

Expose and add default Json serialization to ToolConfig #778

Closed realdavidvega closed 2 weeks ago

realdavidvega commented 3 weeks ago

Background

Currently, tool input decoding and output encoding is using the global Json serialization configuration.

This configuration, by default, uses ignoreUnknownKeys = true in Json serialization, which is useful when the provider, like OpenAI, returns unknown keys that are not supported yet. However, on tools, it provokes swallowing decoding errors of non-strict tool inputs, and can lead to unexpected results. When not swallowing this error, the assistant can recover from this kind of problems by trying again.

Also, by default it uses explicitNulls = false, so the assistant sometimes interprets that the absence of the key means that there is not value. This leads to the assistant not calling a Tool, which not happens so often when we pass a pattern of { key: null }, so the assistant interprets that it has to call a Tool to recover that information.

This PR

In any case, this PR makes the tool decoding and encoding to use a separate Json serialization configuration object, which can be specified per tool, and is by default Json.Default. Allowing to override the Json configuration per tool based on the consumer needs.

In future, we could improve the input encoding by adding the strict: true parameter per tool, as supported in OpenAI.