GenAi - interactions with cached content support?

Gemini API
, ,
1

hey,

we are have a client that processes big file using genai context caching, and it works great.

I was looking into a solution to saving conversation and loading back into them.

genai has a “Interactions API” witch is perfect for us as it handles most of the work and hands out keys to load back into conversations.

the thing is it doesn’t support context caches, i was guessing because context cache has a ttl and is part of a different service,

is there a way to use interactions with context? maby store to convo without the context and load a new cache with the interaction?

    def _create_first_interaction(self, cached_name, textInput) -> str:
        interaction = self.client.interactions.create(
            model=self.gemini_model,
            store=True,
            input=textInput,
            generation_config=types.GenerateContentConfig(
                thinking_config=types.ThinkingConfig(thinking_budget=-1),
                cached_content=cached_name,
                stop_sequences=["[DONE]"],
                temperature=0.0
            ),
        )
        debug("Created interaction: " + interaction)
        self.interaction_id = interaction.get('name')
        return interaction

any other ideas?? without interactions?

2

The syntax above is inaccurate, it seems to be mix & matching generate content syntax with interactions API syntax, you don’t need types.GenerateContentConfig, types.ThinkingConfig etc with interactions API.

See this doc for right syntax: https://ai.google.dev/gemini-api/docs/interactions?ua=chat#stateful-conversation

It does not support explicit context caching API but using stateful mode should increase your chance of cache hits and perhaps will solve your problem with regards to big files regardless?

3

stateful mode should increase your chance of cache hits

I didn’t understand what you meant?

the files are loaded at the start of the session - they are loaded into a context cache…

I think i will skip the interactions and store the history manually in a json, and re init the context with each session.

looking over the docs, why history isn’t also loaded into the cache? like the system prompt?