AI-based order recognition module

Abstract

In this work, presented is a technical description of a software module for automating customer order processing in an industrial company. Analyzed is the finding that manual processing of unstructured orders takes twenty to forty minutes per document and leads to errors when matching against a catalog of four thousand items. Investigated is the microservice architecture based on FastAPI, comprising a RAG service, telemetry collector, and an embedded widget for the Bitrix24 CRM system. Revealed is that hybrid search combining semantic vector representations and a lexical inverted index provides high accuracy in matching customer descriptions against catalog nomenclature. Studied is the application of a three-level text extraction strategy for documents in PDF, Excel, CSV, and ZIP formats with Tesseract OCR fallback. Determined is that text chunking with overlap and a multi-level LLM API invocation strategy enable reliable extraction of product items from orders of arbitrary volume. Established is that fine-tuning the E5-base model on a domain-specific corpus using triplet loss with hard negatives improves semantic search quality. Formed is an iterative accuracy improvement mechanism through a closed-loop telemetry collection and manager feedback system. Proposed is a two-stage nomenclature matching scheme with gate optimization and caching to accelerate processing of typical queries. Developed is integration with Bitrix24 and 1C systems automating the complete cycle from file upload to order creation in the accounting system. Substantiated is the advantage of the proposed approach over manual processing and classical full-text search in both speed and result quality. Presented is a description of validation results on real customer data confirming a three-to-four-fold reduction in order processing time.