JavaScript · Artificial Intelligence · Full Stack

A 200-Line AI Diary That Actually Understands What 'Proud' Means

By dzhd · Jul 27, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The vector-database-plus-LLM pattern shown here is the standard infrastructure for nearly every AI agent product today. Understanding how to wire Milvus, embeddings, and an LLM into a retrieval loop — and how to keep the model from hallucinating when context is thin — is a prerequisite for building agents that actually remember and reason over user data.

Summary

A complete, from-scratch build of an AI diary assistant that can answer questions like "What made me proud recently?" by understanding meaning rather than matching keywords. The system stores diary entries as 1024-dimensional vectors in a Milvus collection, retrieves semantically similar entries using cosine similarity, and feeds them as context to a Qwen LLM for empathetic, first-person answers. The entire pipeline — embedding, indexing, search, and generation — runs on Zilliz Cloud and Alibaba Cloud DashScope, both accessed through a single OpenAI-compatible interface.

The walkthrough covers the full stack: why IVF_FLAT indexing avoids O(n) brute-force scans, how to design a Collection schema with array-typed tags, and the prompt engineering details that keep the LLM faithful to retrieved documents. A key architectural point is that switching LLM or embedding providers requires only a `baseURL` change, since DashScope, DeepSeek, and others all implement the OpenAI API protocol.

Takeaways

— Cosine similarity on text embeddings finds semantically related content even when no words overlap, as with "proud" matching "sense of accomplishment."

— IVF_FLAT indexing clusters vectors with K-Means so queries only scan a few nearby clusters, dropping search complexity from O(n) to roughly O(N + M).

— A Milvus Collection schema supports an Array data type for tags, removing the need for a separate join table.

— Setting the LLM temperature to 0.1 in a RAG pipeline keeps the model faithful to retrieved documents instead of improvising.

— Switching between LLM or embedding providers requires only a `baseURL` change because DashScope, DeepSeek, and others all implement the OpenAI API protocol.

— `loadCollection` must be called before searching; Milvus keeps data on disk until explicitly loaded into memory.

— Using `Promise.all` for batch embedding generation is fast but can trigger rate limits; production code needs a concurrency limiter like `p-limit`.

Conclusions

The OpenAI API has become a protocol-level standard, not just a product. Multiple Chinese LLM vendors implement it, which means a single `@langchain/openai` dependency can target entirely different model backends with a one-line config change.

The prompt's explicit fallback instruction — "If there is no relevant information, gently inform the user" — is a lightweight but effective hallucination guardrail that many RAG demos omit.

Storing tags as a native Array type in Milvus simplifies reads but shifts filtering complexity to scalar filtering, which is a different query pattern than SQL WHERE clauses. Developers accustomed to relational joins may need to rethink their data model.

Concepts & terms

IVF_FLAT

An inverted-file index that clusters vectors using K-Means. At query time, it compares the query vector only against the nearest few cluster centroids, then performs exact similarity search within those clusters. This reduces search scope from the entire dataset to a small fraction, achieving millisecond latency on millions of vectors.

Cosine Similarity

A measure of similarity between two vectors based on the cosine of the angle between them. Values range from -1 (opposite) to 1 (identical direction). It is preferred for text embeddings because it ignores vector magnitude and focuses on directional similarity, so "hiking" and "mountaineering" score high even if their vector lengths differ.

RAG (Retrieval-Augmented Generation)

A pattern where a user query is embedded and used to retrieve relevant documents from a vector store; those documents are then injected into the LLM's prompt as context before generation. This grounds the model's answer in specific, retrieved data rather than relying solely on its training knowledge.

Scalar Filtering

In Milvus, the ability to filter search results by non-vector fields (e.g., tags, dates, moods) alongside a vector similarity search. It replaces traditional SQL WHERE clauses for metadata constraints within a vector query.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗