From f431c18efcde4f100194cdf142dfdc16b3811d3c Mon Sep 17 00:00:00 2001 From: Raghotham Murthy Date: Thu, 11 Jul 2024 01:32:17 -0700 Subject: [PATCH] added more docs --- source/api_definitions.py | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+) diff --git a/source/api_definitions.py b/source/api_definitions.py index 84e0954f7..e2a7150be 100644 --- a/source/api_definitions.py +++ b/source/api_definitions.py @@ -237,6 +237,28 @@ class AgenticSystemExecuteResponseStreamChunk: class AgenticSystem(Protocol): + """The Llama 3 models released by Meta in July should not just be seen as a model, + but really as a system starting the transition towards an entity capable of performing + "agentic" tasks. By that we mean the following specific capabilities: + 1. Ability to act as the central planner -- break a task down and perform multi-step reasoning. + 2. Ability to perceive multimodal inputs -- text, images, files and eventually speech and video in later iterations. + 3. Ability to use tools - + a. built-in: the model has built-in knowledge of tools like search or code interpreter + b. zero-shot: the model can learn to call tools using previously unseen, in-context tool definitions + + Multi-step tool-use concretely helps address many common problems with LLMs that users may + face: + 1. Finding accurate and up-to-date information. LLMs are limited to training data and knowledge cut off date. + 2. Current LLMs are limited in their understanding and reasoning abilities for solving more complex math problems, processing and analyzing data. Tools like code-execution or APIs like Wolfram can help bridge the gap. + 3. Users may need help with a task that requires multiple tools to execute or a task that has multiple steps (e.g., graph plotting, etc.) + 4. Our current LLMs are not able to generate other modalities (images, voice, video) directly. + +Finally, we want the underlying LLM to remain broadly steerable and adaptable to use cases which +need varying levels of safety protection. To enable this, we want to shift safety into a two-tiered +system: + 1. a set of "always on" safety checks are always performed at the model level, and + 2. a set of configurable safety checks which can be run at the overall system level. +""" @webmethod(route="/agentic_system/create") def create_agentic_system(