Why Most RAG Projects Fail and How Tencent’s LeXiang AI Assistant Overcomes Them

1. The booming RAG community

Since ChatGPT sparked interest in intelligent Q&A, many enterprises have tried to build internal "AI experts" using open‑source LLM + RAG pipelines. While the concept is attractive, real‑world deployments quickly encounter high hardware costs, ongoing maintenance burdens, and performance bottlenecks.

A typical case study: a mid‑size company assembled a team, connected a vector database to an open‑source LLM, and launched a RAG service. Within three months the team was overwhelmed—one engineer spent all his time fixing hallucinations, a data engineer handled endless data cleaning, and another struggled with latency and security incidents, while the budget exploded more than fivefold.

2. Rationalizing the approach

Because building a RAG system from scratch requires continuous investment in infrastructure, data ingestion, security, and operations, many enterprises find it more sensible to adopt a mature, cloud‑based product.

2.1 Stable knowledge‑management core

LeXiang AI Assistant treats knowledge quality as the foundation: "Garbage in, garbage out". The product leverages Tencent’s long‑standing expertise in knowledge classification, permission control, and document parsing to ensure that only well‑structured, high‑quality information reaches the LLM, thereby improving answer accuracy.

2.2 Rich industry experience

With over 300,000 registered enterprises and 100 % industry coverage, LeXiang continuously incorporates real‑world customer feedback to fine‑tune its models, resulting in higher answer quality and better product performance.

2.3 Efficient resource management

The assistant provides automated monitoring, fault detection, and intelligent load balancing on Tencent Cloud, allowing elastic scaling of compute resources to match fluctuating traffic while controlling costs.

2.4 Advanced model base and training stack

LeXiang uses Tencent’s proprietary Mixtral large‑model foundation, which excels in content creation, logical reasoning, and multi‑turn dialogue. The Mixtral One‑Stop Platform automates the entire workflow from model training to deployment, accelerating feature rollout.

3. Future development directions

3.1 Modality diversification

Enterprise documents often contain images, diagrams, and PDFs that pure‑text models cannot process. Two practical routes are discussed:

Convert images to text using OCR for simple cases (e.g., scanned pages).

For structured graphics (flowcharts, architecture diagrams), employ a small model to translate them into Markdown‑style descriptions, then feed the result to the LLM.

Long‑term, a fully visual RAG pipeline ("Visual RAG" or OCR‑free) could enable image‑based retrieval and Q&A.

3.2 Knowledge span

When the LLM’s context window is limited, knowledge retrieval remains essential. LeXiang supports a 256k token window (≈200 k characters), allowing it to handle long documents while still relying on retrieval for concise, relevant snippets.

3.3 Reasoning complexity

Beyond simple fact lookup, enterprise queries often require multi‑step reasoning, synthesis across documents, or pattern extraction from historical cases. The article cites research that classifies queries into four levels—from explicit facts to hidden rationales—and argues that future RAG systems should evolve toward agent‑style architectures that combine tool usage, planning, and self‑reflection.

4. Summary

The three dimensions—modality, knowledge span, and reasoning complexity—reflect Tencent’s understanding of enterprise knowledge‑Q&A needs and guide the roadmap of LeXiang AI Assistant. While the product already offers robust knowledge management, extensive industry data, scalable resources, and a powerful LLM foundation, ongoing work will focus on multimodal support, longer context handling, and agent‑based reasoning to further close the gap between naive RAG and truly intelligent assistants.