Feature requests

[creatorrr]

Please upvote or comment with features that you need. We'd love to prioritize them. 🤖

[creatorrr]

olap queries GET /stats

[creatorrr]

dump everything as a csv GET /export

[creatorrr]

• jinja template rendering in instructions (with access to user.*, session.*, agent.*)

[creatorrr]

This was shipped as part of 0.3.2

[philipbalbas]

add multiple users to a session

[creatorrr]

support for replicate / other model providers

[alt-glitch]

import previous/other sessions from different providers by adding in ChatML formatted history

[ijindal1]

expand on RAG and the ability to store "knowledge" in vector DBs

• for eg if an organization makes a lot of agents, then they can access the same knowledge base for all of them. (?)

[ijindal1]

automatic logging and monitoring - token usage, no of users, no of chats per user,

[alt-glitch]

function calling feature parity with openai / anthropic.
ability to add the function call result directly to the prompt/session

[alt-glitch]

either decide to use litellm for function calls, or extend the capabilities in /agents-api/agents_api/routers/sessions/session.py

[creatorrr]

Support for windows and other docker-free environments

[alt-glitch]

Tools to be scoped to users or sessions instead of just agents.

[creatorrr]

Is there a strong use case for this added complexity? We could just have tools on agents but metadata or arguments derived from a user, no?

[alt-glitch]

Cache Features:

• Add use_cache flag/s in session.chat to force use cache even if init_context is different. So instead of init_context being used, ChatML messages individually are hashed, checked and retrieved.

[creatorrr]

what do you mean by this @alt-glitch ?

[whiterabbit1983]

I am wondering if we should implement LLM response cache expiration?

[creatorrr]

Task steps for sleeping and waiting for input.

- sleep: "{{ num }} days"  # plain english time duration like '1 day', '2 hours' etc. must be in the future and parseable by moment.js etc
- wait_for: "input"  # some condition to wait for, right now just "input" which means that it'll wait like it waits for tool call response

[alt-glitch]

Have a create_many function or add support for adding a list of docs. Currently only a single doc is accepted