mia-code 0.2.0 → 0.3.0

package/contextual_research/PDE-generalization--caefee82-efb1-4dbb-8733-691b01581464--260130/sources/footnote_1_28_guardrails_multi_agent.md

@@ -0,0 +1,260 @@
+MARCO: Multi-Agent Real-time Chat Orchestration
+
+Anubhav Shrimal1, Stanley Kanagaraj1, Kriti Biswas1,
+Swarnalatha Raghuraman1, Anish Nediyanchath1,
+Yi Zhang2 and Promod Yenigalla1
+1Retail Business Services, Amazon
+2AWS Bedrock, Amazon
+{shrimaa, kstanly, kritibw, rswarnal, anishned,
+yizhngn, promy}@amazon.com
+
+4
+2
+0
+2
+
+t
+c
+O
+9
+2
+
+]
+I
+
+A
+.
+s
+c
+[
+
+1
+v
+4
+8
+7
+1
+2
+.
+0
+1
+4
+2
+:
+v
+i
+X
+r
+a
+
+Abstract
+
+Large language model advancements have en-
+abled the development of multi-agent frame-
+works to tackle complex, real-world problems
+such as to automate tasks that require interac-
+tions with diverse tools, reasoning, and human
+collaboration. We present MARCO, a Multi-Agent Real-time Chat Orchestration framework
+for automating tasks using LLMs. MARCO ad-
+dresses key challenges in utilizing LLMs for
+complex, multi-step task execution. It incorpo-
+rates robust guardrails to steer LLM behavior,
+validate outputs, and recover from errors that
+stem from inconsistent output formatting,
+function and parameter hallucination, and lack
+of domain knowledge. Through extensive experi-
+ments we demonstrate MARCO’s superior per-
+formance with 94.48% and 92.74% accuracy
+on task execution for Digital Restaurant Ser-
+vice Platform conversations and Retail conver-
+sations datasets respectively along with 44.91%
+improved latency and 33.71% cost reduction.
+We also report effects of guardrails in perfor-
+mance gain along with comparisons of various
+LLM models, both open-source and proprietary.
+The modular and generic design of MARCO
+allows it to be adapted for automating tasks
+across domains and to execute complex use-
+cases through multi-turn interactions.
+
+1
+
+Introduction
+
+Advancements in LLM technology has led to a
+lot of interest in applying Agents framework to
+realise solutions which require complex interac-
+tions with the environment including planning,
+tools usage, reasoning, interaction with humans.
+Recent works (Wang et al., 2024; Huang et al.,
+2024) demonstrate potential of LLMs for creating
+autonomous Agents while there are numerous chal-
+lenges to overcome and provide a seamless experi-
+ence for end users who interact with the system at a
+daily basis. LLMs are probabilistic next token pre-
+diction systems and by design, non-deterministic
+
+which can introduce inconsistencies in the output
+generation that can prove challenging for features
+like function calling, parameter value grounding,
+etc. There are also challenges of domain specific
+knowledge which can be an advantage and dis-
+advantage at the same time. LLMs have biases
+inherent in them which can lead to hallucinations,
+at the same time it may not have the right internal
+domain specific context which needs to be provided
+to get the expected results from an LLM.
+
+We present our work on building a real time con-
+versational task automation assistant framework
+with the following emphasis, (1) Multi-turn In-
+terface for, (a) User conversation to execute tasks
+(b) Executing tools with deterministic graphs pro-
+viding status updates, intermediate results and re-
+quests to fetch additional inputs or clarify from
+user. (2) Controllable Agents using a symbolic
+plan expressed in natural language task execution
+procedure (TEP) to guide the agents through the
+conversation and steps required to solve the task
+(3) Shared Hybrid Memory structure, with Long
+term memory shared across agents which stores
+complete context information with Agent TEPs,
+tool updates, dynamic information and conversa-
+tion turns. (4) Guardrails for ensuring correctness
+of tool invocations, recover for common LLM error
+conditions using reflection and to ensure general
+safety of the system. (5) Evaluation mechanism
+for different aspects and tasks of a multi-agent sys-
+tem.
+
+This is demonstrated in the context of task au-
+tomation assistant which supports adding usecase
+tasks to provide users a conversational interface
+where they can perform their intended actions, mak-
+ing it easier for them to refer to informational doc-
+uments, interact with multiple tools, perform ac-
+tions on them while unifying their interfaces. We
+provide detailed comparison across multiple foun-
+dational LLMs as backbone for our assistant tasks
+like Claude Family models (Anthropic, 2024), Mis-
+
+tral Family models (Jiang et al., 2023, 2024) and
+Llama-3-8B (AI@Meta, 2024) on Digital Restau-
+rant Service Platform (DRSP) conversations and
+Retail conversations (Retail-Conv) datasets.
+
+2 Related Work
+
+Improvements to LLM technology through the re-
+lease of foundational LLMs like GPT-4 (OpenAI
+et al., 2024), Claude (Anthropic, 2024) and Mix-
+tral (Jiang et al., 2024) has led to a flurry of research
+around autonomous agents and frameworks (Wang
+et al., 2024; Huang et al., 2024). Zero shot Chain-
+of-Thought (COT) reasoning (Kojima et al., 2023)
+allows LLMs to perform task reasoning by mak-
+ing it think step by step. LLMs can invoke exter-
+nal tools based on natural language instructions.
+HuggingGPT (Shen et al., 2023b) can coin series
+of model invocations to achieve complex tasks
+mentioned by the user. Toolformer (Schick et al.,
+2023) demonstrates how LLMs can be used as ex-
+ternal tools through API invocations selecting the
+right arguments to be passed from few examples
+and textual instructions. Agents framework (Zhou
+et al., 2023) discuss using natural language sym-
+bolic plans called (Standard Operating Procedures)
+SOPs which define transition rules between states
+as the agent encounters different situations to pro-
+vide more control over agent behavior along with
+memory to store relevant state information within
+the prompt (Fischer, 2023; Rana et al., 2023) or
+long term context externally (Zhu et al., 2023; Park
+et al., 2023). Amazon Bedrock Agents 1 provide
+interface to quickly build, configure and deploy au-
+tonomous agents into business applications leverag-
+ing the strength of foundational models, while the
+framework abstracts the Agent prompt, memory,
+security and API invocations. LangGraph 2 is an ex-
+tension of LangChain which facilitates the creation
+of stateful, multi-actor applications using large lan-
+guage models (LLMs) by adding cycles and per-
+sistence to LLM applications thus enhancing their
+Agentic behavior. It coordinates and checkpoints
+multiple chains (or actors) across cyclic computa-
+tional steps. While these frameworks present novel
+ways for LLMs to act in a desired behaviour, they
+often have accuracy-latency trade-off where to im-
+prove on the accuracy the system latency increases
+due to multi-step planning and thinking (Yao et al.,
+2023; Wei et al., 2023). Our proposed solution,
+
+1Amazon Bedrock Agents User Guide
+2LangGraph library
+
+MARCO, not only interacts with user in a multi-turn fashion but also has multi-turn conversation
+with deterministic multi-step functions which com-
+prises of pre-determined business logic or task ex-
+ecution procedure (TEP) requiring agents only at
+intelligent intervention related steps. Along with
+the usecase TEPs, multi-step functions and robust
+guardrails to steer LLM behaviour, MARCO is
+able to perform complex tasks with high accuracy
+in less time as detailed in subsequent sections.
+
+3 MARCO: Multi-Agent Real-time Chat
+
+Orchestration
+
+In this section, we discuss our approach for
+MARCO. Section 3.1 formulates the problem state-
+ment in terms of Task Automation via real-time
+chat, followed by components of MARCO in sec-
+tion 3.2 and the evaluation methods on performance
+and latency for MARCO in section 3.3.
+
+3.1 Problem Statement
+
+Given an user (Actor), who wishes to perform a
+task with intent I ∈ {OOD, Info, Action}; where
+Out-Of-Domain (OOD) intent is defined as any
+user query which is not in scope of the system
+such as malicious query to jailbreak (Shen et al.,
+2023a; Rao et al., 2024) the system, foul lan-
+guage or unsupported requests, “Info” intent is
+defined as getting information from predefined
+data-sources and indexed documents (Dindex), and
+“Action” intent is defined as a performing a use-
+case related task (ux) which involves following a
+series of instructions/steps (Task Execution Pro-
+cedure, T EPx) defined for the usecase and ac-
+cordingly invoking the right set of tools/functions
+(F x
+2 , ..., F x
+n }) with the identified re-
+quired parameters (P x
+∗ = {PF x
+1 , F x
+2 , ..., F x
+n }, PF x
+∗ = {PF x
+1 , ..., PF x
+n }
+for each function respectively. The objective for a
+task automation system is to, (1) interpret the user
+intent I for each query, (2) identify the relevant
+usecase ux, (3) understand the steps mentioned in
+its T EPx, (4) accordingly call the right sequence
+of tools F x
+∗ with required parameters P x
+∗ , (5) cor-
+relate T EPx, tool responses and requirements and
+conversation context to communicate back with the
+user, and (6) be fast and responsive for a real-time
+chat.
+
+An example scenario is shown in figure 1 where
+User first asks “The sale of certain item is go-
+ing down in my restaurant. Can you please help
+me find out why?”, i.e.
+