Vector Search

混合检索（Hybrid Search）将向量检索与BM25关键词检索的结果通过RRF（倒数排名融合）等算法合并，兼顾语义理解与精确匹配

The core mechanism of vector search is nearest-neighbor search: as long as the vector store is non-empty, it will always return a result, and there is no built-in "no result" branch.

Executing Cosmos DB vector queries usually requires a vector index (such as a DiskANN index) to ensure performance, and vector queries consume relatively many RUs.

Vector search is one of the query capabilities supported by Cosmos DB, used to find the semantically most similar results in vectorized data; Cosmos DB stores vectors in an array property of the document and provides the VectorDistance function to compute distance in queries.

For the E4012 query, BM25 ranks the correct page 1st (score 1/61) and vector search ranks it 14th (1/74); after RRF fusion the correct page scores about 0.030 while the E4021 page scores about 0.016, nearly doubling the correct page's score.

On a self-built test set (120 real queries, 38 of which contain identifiers), pure vector retrieval hit 21/38 identifier queries, pure BM25 (with an identifier-preserving tokenizer) hit 36/38, and hybrid + RRF hit all 38.

If the queries are all fuzzy natural-language questions containing no IDs or terminology, pure vector retrieval may suffice; if the knowledge base is small enough to fit directly into the prompt (Anthropic suggests under 200,000 tokens), you can also try putting it in directly first.

Pure vector search (embedding retrieval) tends to lose rare but precise tokens such as error codes, ticket IDs, SKUs, and function names, because embeddings blur rare tokens when compressing text into a single point.

In scenarios requiring exact keywords (such as the error code AUTH_403), vector search ranking may not be ideal.

Vector embeddings excel at text similarity retrieval but cannot effectively handle structured graph constraints, such as simultaneously satisfying multiple conditions like geographic distance, available ICU beds, on-duty specialists, and allergy contraindications.

Vector representations are compressed semantic summaries that lose details such as exact identifiers, negations and numerical boundaries, and may also rank texts that are semantically similar but opposite in conclusion ahead of others.

Vector search maps queries and documents into dense vectors in a high-dimensional space, measuring semantic closeness via cosine similarity or dot product.

Baruzzo's argument is not against vector search itself, but about the order of doing things: first establish a measurable baseline, and let error types drive technical decisions.

Vector search compares semantic similarity and is inherently poor at exact matching (such as numbers, dates, and codes); directed queries are a weakness of vector search.

Without a baseline, adding vector search can only be called a "change", not an "improvement"; any newly added technical solution must be able to beat the baseline to count as an improvement.

When the evaluation set shows that vocabulary-mismatch errors dominate, hybrid search (keyword plus vector) has sufficient justification, and the benefit can be quantified.

Semantic retrieval (vector search over stored knowledge) and state recovery (loading a checkpointed workflow) solve different problems; vector search can surface relevant context, but should never be the sole mechanism for restoring the exact state of an in-progress task.

Vector search is a technique that encodes unstructured data such as text and images into high-dimensional vectors (embeddings), then uses approximate nearest neighbor (ANN) algorithms to retrieve semantically similar items; common implementations include FAISS, Milvus, Pinecone and pgvector.

Compared with keyword inverted indexes, vector search can capture expressions that are semantically identical but literally different, but it requires an additional embedding service and index maintenance costs.

Vector search alone is powerless for exact matches (such as product numbers, error IDs, version numbers and names); when a user searches for 'error E4021', vector search loses the exact match.