How Taotian Built a Massive Product Knowledge Base with Milvus for AI Shopping Guides

From Natural Language Questions to Usable Product Knowledge

When consumers ask questions like "What skin type is this foundation suitable for?" or "What's the difference between these two SKUs?", the AI shopping guide must retrieve reliable knowledge for the current store and product, then compose an answer. Product details, merchant FAQs, manuals, and platform-generated knowledge all need to be searchable and keep pace with product information changes.

Taotian built a unified product knowledge base system during its e-commerce AI transformation and chose Milvus to carry vector retrieval for product knowledge and FAQs, providing knowledge support for store AI shopping guides. The guides have been deployed in key industries such as beauty and consumer electronics, covering product recommendation, knowledge Q&A, usage guidance, and after-sales explanation.

The core requirements and how Milvus addresses them are summarized below:

FAQ, product wiki, and manuals jointly serve Q&A → Organize knowledge from different sources into retrievable content → Milvus stores knowledge vectors and provides semantic recall for FAQs and wikis.

More merchants, products, and knowledge continuously onboard → Support vector scale growth and online queries → Milvus handles large-scale vector storage, indexing, and retrieval.

Similar products and different SKUs easily confused → Add business scope constraints during semantic retrieval → Milvus combines metadata filtering with sellerId, itemId, skuId.

Merchants correct content, algorithms continuously generate knowledge → Support continuous writes and updates → Milvus syncs approved new knowledge into the retrieval pipeline.

Taotian previously used cloud Tair (Redis-compatible) for vector search in other products. For this knowledge base, selection focused on larger-scale FAQ and wiki data, plus store/product/SKU filtering and continuous update needs. Milvus proved a better fit for such knowledge base workloads.

How Milvus Integrates into the AI Shopping Guide Pipeline

Unifying Platform and Merchant Knowledge

The platform side generates product wikis from product details, SKU info, images, parameters, reviews, "Ask Everyone" data, and transaction data using algorithms and LLMs. Wikis cover selling points, usage instructions, specifications, applicable scenarios, fulfillment and service descriptions.

The merchant side supports FAQ creation, editing, deletion, and Excel bulk import, plus upload of manuals, tutorials, certificates, PDFs, PPTs, and annotation corrections to platform-generated wikis.

Before ingestion, content goes through automated evaluation, merchant correction, staff review, version management, conflict detection, and quality scoring. Only approved content enters the retrieval pipeline.

Merchants annotate and correct product knowledge; after approval it syncs to the retrieval pipeline.

FAQ management supports continuous merchant supplementation and maintenance of product Q&A.

Layered Storage Architecture

The system uses a layered storage approach where each component has a clear responsibility:

MySQL — Stores FAQ relational data, document metadata, audit status, annotation records. Responsible for business management and knowledge governance.

OSS — Stores original manuals, images, and other document files. Responsible for raw file preservation.

Milvus — Stores FAQ and product wiki vectors plus related business metadata. Responsible for semantic retrieval and business condition filtering.

Before entering Milvus, knowledge content undergoes cleaning, chunking, and vectorization. FAQs and wikis are organized into separate retrieval paths; the business system can invoke the appropriate path based on question type and knowledge category.

System architecture: Milvus sits in the RAG retrieval layer, connecting knowledge governance with downstream AI applications.

Business-Scoped Retrieval and LLM Answer Generation

After a consumer asks a question, the system vectorizes the query and identifies business constraints such as sellerId, itemId, and skuId. It then retrieves relevant FAQs and wiki knowledge from Milvus. The recalled content, together with product metadata, historical memory, and session context, is fed to an LLM to generate an answer, with verification and feedback mechanisms for continuous optimization.

For example, for the question "What's the difference between the maternal-infant mode and premium mode of this refrigerator?", FAQs provide the merchant-maintained standard explanation, while wikis supplement functional details from manuals. Restricting retrieval to the specific store, product, and SKU reduces the risk of citing knowledge from other products.

When merchants add FAQs, correct product descriptions, or the platform regenerates wikis, the content passes through the governance flow and vector sync, then becomes available for retrieval. The actual effective time depends on the full process of review, sync, and retrieval visibility.

Value Delivered by This Practice

More complete knowledge coverage. FAQs handle high-frequency inquiries; product wikis supplement tutorials, parameter differences, applicability conditions, and fine-grained SKU information, providing more reference content for complex and long-tail questions.

Recall closer to business context. Vector similarity works together with store, product, and SKU conditions during retrieval, helping the AI shopping guide find answers within a well-defined business scope.

Merchant-maintained knowledge enters real Q&A. FAQs, manuals, and annotation corrections flow into the retrieval pipeline after approval, so merchant knowledge maintenance directly serves consumer inquiries.

Knowledge reuse for more AI applications. The same product knowledge can further serve detail-page AI assistants, customer service, and search shopping guides, and leaves room for future expansion to brand, category, and SPU-level knowledge.

Why Alibaba Cloud Milvus Is Recommended for Similar Scenarios

For teams taking a knowledge base into production, retrieval effectiveness, resource cost, and long-term operations must be considered. Alibaba Cloud Milvus offers a fully managed service compatible with the open-source Milvus ecosystem, supporting elastic scaling and monitoring/alerting, suitable for teams wanting to leverage the Milvus ecosystem while reducing cluster deployment and daily ops overhead.

Learn about Alibaba Cloud Milvus: https://help.aliyun.com/zh/milvus/

Combining the needs from the Taotian case, three directions are worth further evaluation:

EAGLE Kernel for Online QA Performance

Store AI shopping guides need retrieval integrated into online Q&A recall rate, throughput, and tail latency jointly affect user experience. Alibaba Cloud Milvus's self-developed EAGLE kernel optimizes retrieval performance and resource efficiency for both all-in-memory and memory-disk hybrid scenarios, providing an engine option worth attention for such needs.

Recommended reading: "Alibaba Cloud Milvus Self-Developed Kernel EAGLE Released: Vector Retrieval SOTA, We Decided to Build It Ourselves" — suitable for teams focused on online RAG, vector recall performance, and cost-performance. Benchmarks in the article help understand optimization directions; actual selection should still be validated against your own vector dimensions, filter conditions, recall targets, and concurrency.

Disk Index for Massive Vector Scale

As product knowledge, historical data, and multimodal data continuously accumulate, teams must balance retrieval capability with memory overhead. Alibaba Cloud Milvus's disk index optimizes for massive vector retrieval and also covers million-scale TopK large-candidate-set recall needs.

Recommended reading: "Alibaba Cloud Milvus Disk Index New Tool: Million-Level TopK High-Performance Recall" — suitable for teams concerned with large-scale vector capacity, resource cost, and large-candidate-set recall tasks like training data filtering. Daily shopping guide Q&A usually needs only a limited number of relevant chunks; TopK should be set according to subsequent reranking and generation needs.

Vector Lake for Lance-Format Data

If data already resides in Lance format on OSS, teams can further evaluate Alibaba Cloud Milvus Vector Lake to understand how to provide vector retrieval services on top of existing lake data, reducing duplicate data movement and maintenance of multiple copies.

Recommended reading: "How to Choose Lance Data Retrieval? Of Course Alibaba Cloud Milvus Vector Lake" — suitable for teams already using Lance format or planning lake-based vector retrieval. Raw documents stored in OSS and the Lance data organization used by Vector Lake have different requirements; evaluate the integration path based on existing data formats.

The three capabilities above are recommended directions for peer business selection and future evolution. The Taotian case described the practice of Milvus-powered FAQ and product wiki vector retrieval.

For teams building similar product knowledge bases, it is suggested to start with dual-path recall of FAQs and wikis, adding store/product/SKU filter conditions to validate against real questions, then choose Alibaba Cloud Milvus instance and index solutions based on knowledge scale, concurrency, and cost targets. When larger-scale recall or lake-based data retrieval is needed later, extend along the capabilities outlined above.