Tencent Cloud Big Data Computing Intelligence: Unifying SQL and Multimodal AI Workloads

Background: Four Shifts in Data Platform Computing

The article outlines four fundamental changes as data platforms evolve from tools for human analysts to compute foundations for AI agents:

Data assets expand to multimodal: Contracts, reports, recordings, videos, images, logs, embeddings, model features, and agent memories become queryable, callable assets.

AI agents become high-frequency users: Unlike low-frequency human SQL queries, agents run 7x24 continuous task chains — generating queries, triggering retrieval, invoking inference, submitting training, and writing results back.

Workloads become composite chains: Real business flows combine multi-source reads, document parsing, video frame extraction, embedding, hybrid vector/full-text search, reranking, LLM inference, SQL validation, traditional ML prediction, report generation, and knowledge-base write-back.

Competitiveness shifts from computing efficiency to agent efficiency: Metrics now include average task cost, chain success rate, RAG recall accuracy, inference latency, GPU utilization, data lineage traceability, and auto-retrospective capability.

Computing Intelligence: A Fusion Compute System for Agent-Ready Platforms

Computing intelligence is defined as a unified compute layer that supports structured SQL, multimodal processing, traditional ML training/inference, RAG, and agent orchestration — with intelligent scheduling across heterogeneous CPU/GPU resources. It bridges the gap from AI-Ready (can agents access data?) to Agent-Ready (can agents continuously, reliably, and cost-effectively invoke data and compute to achieve business goals).

Product Implementations

Meson: High-Performance Structured Compute Engine

Meson optimizes Spark SQL with vectorized execution, pipeline model, columnar shuffle, remote shuffle, and SSD cache. In 2026 it adds three capabilities aligned with agent-ready demands:

Spark Rapids offloads shuffle, join, and other heavy operators to GPU via time-sliced scheduling of Data+AI platform GPU resources.

Incremental compute avoids full recomputation for high-frequency agent analyses.

Full Spark 4.0 support ensures compatibility with the latest ecosystem.

Results: On 1 TB TPC-DS storage-compute separation, overall performance improves 3.6× (compute-intensive up to 5×), CPU utilization drops from 80% to 40%, serving 100+ customers.

Xpark: Distributed Multimodal Compute Engine

Xpark extends Ray Data with Xpark DataSet and bundles 50+ operators for text deduplication, summarization, video splitting, image augmentation, and common pipelines. It lets developers process text, images, audio, and video with unified Dataset/DataFrame semantics instead of stitching PySpark, OpenCV, Librosa, Python scripts, model services, and vector stores.

Performance highlights:

Text MinHash Dedup reaches 8× Data-Juicer throughput.

Distributed Exact Substring Dedup runs 50× faster than open-source equivalents.

Embedding supports hybrid GPU/CPU inference; CLIP async inference yields ~1.5× speedup .

Reliability features: operator-level retry, task redistribution, skip strategy, lineage reconstruction, and task-level checkpointing reduce full re-runs from single-point failures or data anomalies.

Use cases: PB-scale LLM pre-training corpus cleaning (dedup, quality filtering, PII masking, tokenization), multimodal model prep (video splitting, audio extraction, image quality filtering, image-text similarity, video captioning, safety detection), and AI generation apps (sentiment analysis, translation, summarization, multimodal retrieval).

TCHouse-X AI Function: SQL as the Gateway to AI

TCHouse-X (Tencent Cloud's self-developed cloud data warehouse) integrates with Xpark to expose AI Functions via standard SQL semantics. Analysts can call distributed multimodal compute and LLM/ML training/inference directly in SQL: AI kmeans for customer clustering AI frequency for purchase frequency CREATE MODEL to define a KMeans model, then SELECT predict(...) to get cluster IDs, purchase frequency, return rates, etc.

This eliminates the old export → Python → model service → dependency management → write-back loop, lowering the adoption barrier for business analysts.

TCRay: Next-Generation AI Workload Infrastructure

Built on open-source Ray, TCRay deepens CPU/GPU unified scheduling for batch, training, inference, RAG, and agent services. Key optimizations target:

Job scheduling, task recovery, resource elasticity, runtime isolation, and large-scale job stability — reducing failure rates and debugging cost for long-running, batch inference, and training jobs.

Ecosystem integration: Ray Data, Ray Train, Ray Serve workflows; training job management, model registry, online serving, offline/online resource sharing, and domestic hardware adaptation.

Reusable templates for intelligent recommendation, RAG, document intelligence, batch inference, and LLM fine-tuning.

ES One-Stop RAG Service: Evaluatable Retrieval-Augmented Generation

The new RAG Effect Evaluation moves RAG from "retrievable" to "verifiable and optimizable." It continuously assesses knowledge bases, embedding models, chunking strategies, hybrid search, and reranking across dimensions:

Recall hit rate

Answer relevance

Factual consistency

Citation traceability

This helps pinpoint "retrieval inaccuracy," "answer hallucination," and "missing context," turning RAG into an observable, tunable, closed-loop enterprise knowledge base. ES serves as the retrieval exit in the computing intelligence layer, collaborating with TCRay for RAG inference and TCLake for storage.

Open Compute Pool & Future Outlook

All engines expose capabilities via SQL, Python, MCP, Skill, and CLI — callable by DataBuddy, WorkBuddy, CodeBuddy, and third-party agent platforms. The platform becomes a shared compute pool for both human analysts and AI agents.

Looking ahead:

Primary users shift from humans to human + agent ; machine-friendly interfaces, controllable permissions, explainable results, auditable calls, and recoverable execution become mandatory.

Compute objects expand from structured tables to multimodal context data (text semantics, image content, video segments, audio emotion, vector similarity, graph relations, model features, agent memory).

Paradigm shifts from single-engine optimization to composite chain optimization — a business task may invoke SQL engine, multimodal engine, vector search, full-text search, model serving, training frameworks, and agent orchestration. Platform competitiveness is measured by end-to-end chain efficiency.

The deep end of Data+AI is not whether an enterprise has plugged in an LLM, but whether it owns a data foundation that lets multimodal data be continuously understood, computed, invoked, and governed by agents. Tencent Cloud's computing intelligence solution aims to unify structured compute, multimodal compute, model inference, RAG retrieval, and agent orchestration into a single governable, observable, and continuously evolvable platform for building enterprise-grade AI-native compute foundations.