Implementations to build the Flare (Forward-Looking Active REtrieval augmented generation) pattern with LangStream

[eolivelli]

This patch includes a set of improvements needed to support the Flare pattern.

Side implementations:

• add support for "loop-over" in the "query"/"query-vector-db" agents
• add support for "loop-over" in the "compute-ai-embeddings" agents
• enhance the error message when you try to access a property on a message that has malformed JSON
• fix "logprobs" field in "ai-text-embeddings", it must be a map and not a JSON string

Flare implementation:

• Look up the main reference here: https://github.com/jzbjyb/FLARE/tree/main
• Add the new "flare-controller" agent
• Add an example application

Notes:
Flare can be used only with OpenAI (/Azure) and only with the text-completion API because you need the "logprobs" feature.

How it works:

• the text-generation agents output a result together with the logprobs (probability of certainty) for all the returned tokens
• the flare-controller finds sub sequences in the answer that are with low confidence
• in case there are log confidence sequences then the flare-controller puts back the message into a loop and requests more documents for retrieval in order to help the LLM

There is some protection mechanism that prevents to run too many iterations on the same message.