Building a Commercial Intelligence Assistant with Baidu's Wenxin Large Model: Methods, Optimizations, and Future Outlook

The presentation introduces the use of Baidu's Wenxin large model to create a commercial intelligence assistant, focusing on how it improves business revenue and user experience for the "AiQiCha" platform.

It first outlines five parts: an introduction to commercial information queries, several model‑based construction modes, advanced techniques, future prospects, and a Q&A session.

Commercial information queries are described with typical scenarios such as partner evaluation, sales outreach, cost control, consumer decisions, and investment analysis, emphasizing the value of large‑scale enterprise data.

The "Copilot" system, deployed in early 2023, achieved notable metrics by March: a 52% increase in dialogue satisfaction, a 54% rise in conversation initiation rate, and a 329% boost in daily lead capture, demonstrating the commercial impact of AI‑driven matching.

Four construction modes are discussed. The first uses Retrieval‑Augmented Generation (RAG), which alone struggles with deep commercial knowledge bases, prompting the need for a proprietary enterprise knowledge graph.

To overcome RAG limitations, a hybrid solution combines a self‑built knowledge base with the large model, performing intent recognition, precise attribute extraction, and SQL generation. Accuracy is improved by injecting table schema information and applying in‑context learning with example queries, raising code generation correctness from ~10% to 70‑80%.

Because large models have limited context windows, a "schema linking" strategy is employed: dynamically extracting only the required tables and fields for a query, reducing prompt size while preserving accuracy.

Advanced applications include self‑reflection for generated graph queries, enabling the model to correct edge direction or node‑edge mismatches, achieving over 90% online accuracy.

Visualization is achieved with Apache ECharts, turning model‑generated data into interactive charts, and the assistant is also explored for enterprise risk analysis.

The future outlook envisions real‑time data analysis in meetings, proactive opportunity detection, and risk assessment, turning the assistant into a productivity‑enhancing partner.

The Q&A covers deployment beyond AiQiCha, prompt engineering, internal usage results, evaluation methods, fine‑tuning versus example injection, and Baidu's Qianfan platform for model fine‑tuning.