Inside Tencent LeYong AI: Solving Enterprise RAG with Knowledge, Engineering & Algorithms

ChatGPT sparked a surge in intelligent Q&A solutions, prompting many enterprises to adopt AI assistants that act as internal "AI experts" for faster knowledge retrieval and utilization. Retrieval‑Augmented Generation (RAG) provides the technical foundation, and Tencent LeYong has built a comprehensive RAG roadmap for its customers.

Knowledge Management Capability

Enterprise AI assistants often include a "knowledge Q&A" feature that allows uploading documents for querying. To capture collective intelligence, a collaborative knowledge‑management platform is essential.

LeYong AI assistant can summarize and answer questions based on site‑wide knowledge, providing reference links.

LeYong evaluates knowledge from three angles: whether the knowledge exists, its quality, and how well it is managed.

1.1 Knowledge Boundary (Is the knowledge present?)

Large models already contain vast world knowledge, but enterprises must decide which proprietary knowledge to feed into the model. A common pitfall is assuming that internal expertise is already known to the model, leading to unanswered queries.

1.2 Knowledge Quality (Is the knowledge high‑quality?)

Uploading massive document collections can suffer from "garbage in, garbage out" if parsing is poor. Complex formats in PDFs, PPTs, and Excel files require robust layout, paragraph, and table extraction, especially for image parsing.

Customers often need two types of image handling:

Macro: Provide concise captions for engineering drawings or schematics so the model can reference the image correctly.

Micro: Convert diagrams, flowcharts, and mind maps to markdown‑style text to preserve information with minimal loss.

LeYong AI assistant provides image‑rich answers and helps users locate relevant pictures quickly.

1.3 Knowledge Permissions (Is the knowledge well‑governed?)

Enterprises have complex permission hierarchies across departments, subsidiaries, and external partners. LeYong inherits Tencent's internal KM platform experience to enforce fine‑grained access control, ensuring users only retrieve documents they are authorized to see.

Engineering Capability

Beyond the model itself, a robust engineering architecture determines the lower bound of RAG performance. The pipeline includes query preprocessing, text splitting, indexing, and model iteration.

2.1 Retrieval

Retrieval is often the weakest link; it involves extensive debugging and optimization of text segmentation, vectorization, keyword indexing, and relevance ranking. Distinguishing between pure Q&A and broader conversational intents is crucial.

2.2 Model Training & Deployment

Fine‑tuning large language models requires selecting an appropriate base (e.g., 7B vs. 70B, MoE vs. dense), preparing high‑quality labeled datasets, running distributed training, and establishing automated evaluation metrics for knowledge‑Q&A scenarios. Tencent’s “Hybrid One‑Stop” platform streamlines model selection, data quality checks, fine‑tuning, evaluation, and deployment with visual tools.

Algorithmic Capability

To achieve production‑grade Q&A, enterprises must advance RAG in three algorithmic directions: multimodal diversity, knowledge span, and reasoning complexity.

3.1 Multimodal Diversity

Beyond OCR‑based image captioning, many industrial images (e.g., schematics, logos) require OCR‑free approaches that embed both text and visual features into a shared semantic space, enabling accurate image retrieval and generation.

3.2 Knowledge Span

Long‑context windows are essential for cross‑document queries and extensive single‑document summarization. Tencent’s Hybrid model supports a 256k token window (~200,000 characters), allowing comprehensive processing of long texts.

Figure from "Retrieval Augmented Generation (RAG) and Beyond" survey.

3.3 Reasoning Complexity

Reasoning levels range from explicit fact retrieval to hidden rationales that require pattern extraction from historical cases. Current RAG systems struggle with the higher levels, highlighting the need for agent‑style architectures and stronger internal reasoning, as exemplified by models like o1.

Overall, Tencent LeYong’s three‑dimensional framework—knowledge management, engineering, and algorithms—captures the challenges and solutions for enterprise RAG, and points toward future directions such as multimodal agents and advanced reasoning capabilities.