How DeepSeek and TiDB AI Are Redefining Data Engines for the Large‑Model Era

DeepSeek’s Transformative Impact

DeepSeek, as a cost‑effective, open‑source large model, is reshaping industry by providing cheap compute, removing monopoly, enabling real‑time business decisions, and supporting diverse integration scenarios such as cloud services, mobile apps, and private deployments.

Its V3 model is inexpensive, while the R1 model offers higher‑quality inference; both are accessible in China, unlike OpenAI.

PingCAP’s AI Product Journey

Since early 2023 PingCAP has built AI‑enhanced database products, beginning with TiDB Cloud Chat2Query (SQL Copilot) powered by OpenAI GPT‑3, which translates natural‑language queries into SQL.

Subsequent improvements include model upgrades (GPT‑4), prompt‑engineering, retrieval‑augmented generation (RAG), multi‑step iteration, and integration of knowledge‑graph‑based RAG (TiDB.AI).

TiDB.AI RAG Framework

TiDB.AI combines vector and graph RAG, supporting keyword, vector, and graph retrieval, and dynamically adjusts workflow complexity based on question difficulty.

It powers applications such as TiDB documentation search, AskTUG, and Slack channel Q&A.

AutoFlow Framework

AutoFlow abstracts knowledge‑graph construction and workflow orchestration, allowing developers to install via pip and build custom knowledge‑graph applications.

It evolves from simple vector RAG to graph RAG and full RAG workflow orchestration.

Unified Data Engine in TiDB

TiDB unifies row, column, vector, and graph storage in a single database, enabling “One Database for All” retrieval capabilities, including vector search for AI‑driven queries.

AI features such as multilingual full‑text search, SQL generation, migration assistance, and operational intelligence (index advisor, anomaly detection) are integrated.

Q&A Highlights

AutoFlow can automatically build knowledge graphs, which are editable for correction.

Vector search resolves mismatched value queries by semantic similarity.

Long documents are chunked before embedding; vectors and chunks can be stored together for efficient retrieval.

Complex OLAP queries benefit from multi‑step generation and agent‑style execution.

Overall, the session demonstrates how large‑model advances, open‑source models like DeepSeek, and TiDB’s AI‑native architecture together form a data‑centric AI ecosystem.