ReAct Agent

Intelligent consultation system for e-commerce

Solution Overview

ReAct Agent is an intelligent dialogue system based on Reasoning + Acting technology, designed to automate customer service in e-commerce. The system combines natural language processing capabilities with real-time access to business data.

Cached responses60-80%
Response time (cache)<0.1 sec
AI cost savingsup to 95%
Request coverage99.9%

Market Positioning

Problems with existing solutions

Typical AI systems operate in one of two extreme scenarios. Either you get a response that has nothing to do with your question - a template reply that doesn't solve the problem - and you can't easily reach a live operator. Or the opposite: at the slightest deviation from the script, there's an instant transfer to an operator, completely devaluing the bot's presence.

Architectural Solution

ReAct Agent implements a fundamentally different approach:

  • Deterministic responses - the system generates answers exclusively based on verified data from the knowledge base. When relevant information is absent - proper escalation instead of generating unreliable content.
  • Intelligent routing - automatic determination of operator involvement need based on dialogue context analysis, before customer frustration occurs.
  • Maximizing autonomous resolution - full utilization of available toolkit before escalation: searching multiple sources, clarifying details, combining data.

Technical Architecture

Agent Model

Unlike pipeline RAG systems where information is provided to the model in a fixed format, ReAct Agent functions as an autonomous researcher. The system independently determines required data and initiates its retrieval.

Example:

When needed, the agent requests the full article from the knowledge base rather than being limited to provided fragments - a capability unavailable in standard RAG systems.

Modular Tool System

The architecture supports unlimited number of tools. The agent autonomously selects the optimal tool depending on the task, for example:

  • Semantic search - meaning-based search accounting for synonyms, typos, and different phrasings
  • Attribute search - filtering and sorting by price, category, characteristics
  • CRM integration - access to order data, statuses, customer history
  • Full-text access - retrieving complete articles from knowledge base

Request Processing Cycle (ReAct)

ReAct (Reasoning + Acting) technology provides an iterative processing workflow:

  1. Reasoning - request analysis and hypothesis formation about necessary actions
  2. Acting - execution of selected tool with corresponding parameters
  3. Observation - result analysis and decision about next step: continue searching or form final response

Security System

Each request passes through a five-level security perimeter:

1
Rate LimitingSpam protection, 10 requests/min
2
Content ModerationAI intent analysis, abuse blocking
3
Semantic CacheSimilar request search, 85% threshold
4
Usage LimitsDaily limits, 100 messages/day
5
AI AgentFull processing

Caching System

Three-level caching architecture provides up to 95% savings on inference costs:

Verified Cache - manually verified responses, 20-30% hit rate.

Semantic Cache - automatic by semantic similarity, 60-80% hit rate.

Exact Match Cache - exact request match, 5-10% hit rate.

Key feature: cache considers dialogue context. Identical requests in different contexts receive corresponding responses from different cache entries.

Semantic Search

The system uses 1024-dimensional vector representations (embeddings) for meaning-based search:

Query vectorization - text conversion to numerical vector, "semantic fingerprint".

Nearest neighbor search - identification of 30 closest documents by cosine distance.

Reranking - re-ranking and selection of 8 most relevant results.

Response generation - forming response based on selected context.

Performance and Economics

Cached response0.05-0.1 sec
Simple question1-2 sec
Question with search2-5 sec
Complex request5-10 sec

Cost of processing 1000 requests: without caching ~$5.70, with 80% hit rate ~$1.14 (80% savings).

Key Advantages

  • Determinism - responses exclusively based on verified data
  • Intelligent escalation - automatic operator necessity detection
  • Modularity - unlimited number of integrable tools
  • Performance - 60-80% of responses in fractions of a second
  • Security - five-level protection system
  • Cost efficiency - up to 95% savings on inference
  • Fault tolerance - automatic recovery and backup models