Redis Vector Search: English for Engineers Building AI Applications

Vector search has moved from research papers into production AI applications, and Redis has become one of the most popular choices for storing and querying embeddings at low latency. Engineers building semantic search, recommendation engines, or retrieval-augmented generation (RAG) pipelines need to discuss index types, query strategies, and schema design with precision. This guide covers the English vocabulary that comes up in design reviews, pull request comments, and team discussions when Redis Vector Search is on the stack.

Key Vocabulary

VSS (Vector Similarity Search) — the Redis capability, part of the RediSearch module, that allows storing high-dimensional vectors alongside regular data fields and querying them by similarity rather than exact match. “We chose VSS over a dedicated vector database because we were already running Redis in production and didn’t want to introduce another stateful service.”

HNSW index — Hierarchical Navigable Small World, an approximate nearest-neighbour algorithm that trades a small amount of recall accuracy for dramatically faster query performance at scale. “We switched from a FLAT index to an HNSW index once the embedding count exceeded half a million — query latency dropped from 80 ms to under 5 ms.”

FT.CREATE — the Redis command used to create a search index, including the schema definition for text, numeric, tag, and vector fields. “The schema in FT.CREATE defines a VECTOR field with the HNSW algorithm and sets the dimension count to 1536 to match the OpenAI Ada embeddings.”

Vector field — a field type within a RediSearch index schema that stores a fixed-dimension floating-point array representing an embedding. “Each product document has a description_embedding vector field populated by passing the product description through the embedding model at index time.”

KNN query — a K-nearest-neighbour query that returns the K most similar vectors to a given query vector, ranked by cosine or Euclidean distance. “The search endpoint runs a KNN query with K set to 10, then re-ranks the results using a lightweight cross-encoder before returning the top three to the client.”

Hybrid search — a query that combines a vector similarity score with traditional filter conditions (tag, numeric range, full-text) so that results must satisfy both semantic relevance and structured criteria. “We implemented hybrid search so users can say ‘find products similar to this description but only in the footwear category and under £50’ — the pre-filter happens before the KNN stage.”

Redis Stack — the distribution of Redis that bundles core Redis with a set of modules including RediSearch (for vector and full-text search), RedisJSON, RedisTimeSeries, and RedisBloom. “Make sure the Docker image you pull is redis/redis-stack, not plain redis — VSS requires the RediSearch module that only ships with Redis Stack.”

Embedding dimension — the number of floating-point values in a vector, which must match exactly between the index schema and every vector stored or queried against it. “We hit a dimension mismatch error in staging because the embedding model was updated from 768 to 1024 dimensions but the index schema wasn’t recreated.”

Useful Phrases

• “I’ll run FT.CREATE with a VECTOR field using the HNSW algorithm, cosine distance metric, and a dimension of 1536 — then we can start ingesting embeddings.”
• “The pre-filter in the hybrid query is reducing the candidate set before the KNN step, which keeps recall high without scanning the full index.”
• “We’re storing the raw JSON document with JSON.SET and keeping the vector in a separate hash field indexed by RediSearch — that way we can update the embedding without rewriting the whole document.”
• “At a million vectors, FLAT search becomes too slow for real-time use; HNSW with EF_RUNTIME tuned to 200 gives us the right balance of speed and recall for our use case.”
• “The FT.SEARCH command returns both the distance score and the document fields — we normalise the cosine distance to a similarity percentage before displaying it in the UI.”

Common Mistakes

Saying “search by vector” instead of “query by similarity”. Non-native speakers sometimes describe a KNN query as “searching by the vector,” which sounds awkward to English-speaking engineers. The standard phrasing is “query by similarity,” “run a similarity search,” or “perform a nearest-neighbour lookup.” Saying “I’m querying by cosine similarity” is clear and natural.

Mixing up “index” as a noun and a verb without the right preposition. Engineers sometimes say “we index the embeddings to Redis” when the correct phrasing is “we index the embeddings in Redis” or “we store and index the embeddings using RediSearch.” The index itself is in (or on) the database, not to it. This distinction matters in written documentation and design documents.

Confusing “dimension” and “dimensionality”. Both words are correct, but dimension is used as a countable noun in practical Redis conversations: “the vector has 1536 dimensions” or “the embedding dimension is 1536.” Dimensionality is more common in academic writing (“the curse of dimensionality”). In daily engineering English, prefer the simpler dimensions or embedding dimension for clarity.

Mastering this vocabulary lets you participate fully in architecture discussions about AI search systems — from initial schema design through to performance tuning and production incident analysis.

Navigating Nuances: A Focus on Clarity for Non-Native Speakers

Redis Vector Search (VSS) is gaining traction in AI application development, and understanding the terminology is crucial – not just for efficiency, but also for clear communication within a technical team. For developers learning professional English, particularly those whose first language isn’t English, the precision of these terms can feel daunting. Let’s consider how to approach this challenge, focusing on building confidence in expressing yourself accurately and understanding complex discussions around vector search.

One common hurdle is the difference between “index” and “algorithm.” Many translations directly use “index” for HNSW (Hierarchical Navigable Small World), which while technically accurate, can obscure the underlying process. Instead of saying “we’re using an HNSW index,” a more effective phrasing – particularly when explaining to a colleague - would be “We’re leveraging the HNSW algorithm to create a fast and scalable vector search solution.” Similarly, discussing FT.CREATE commands requires careful wording. Rather than simply stating “create a vector field,” it’s better to say, “We’re using FT.CREATE to initialize a new vector field containing our embedding data.” This emphasizes the action being performed and avoids jargon that might not be immediately clear. During code reviews, you might see comments like: “Could you elaborate on the rationale behind selecting this specific HNSW index size? A smaller index may result in increased latency during KNN queries.” – a request for clarification framed clearly, avoiding assumptions about the developer’s understanding of the underlying performance trade-offs. Remember that active listening and asking clarifying questions are essential skills when working with technical vocabulary. Don’t hesitate to say, “Could you explain what you mean by ‘optimizing the KNN query parameters’?”

Another area where confusion often arises is in describing the types of queries. “KNN” (K Nearest Neighbors) can feel abstract. Instead of simply stating “we’re performing a KNN query,” consider framing it as: “We’re using a KNN query to identify the most similar vectors within our dataset, based on cosine similarity – allowing us to retrieve items with relevant semantic content.” This adds context and highlights the practical outcome. When writing PR descriptions, precision matters. Instead of saying “Implemented hybrid search,” which is vague, describe it as: “Implemented a hybrid search strategy combining exact phrase matching with vector similarity searches to improve recall for both structured and unstructured data.” This level of detail demonstrates a solid understanding of the system’s architecture and its intended functionality.

Finally, remember that professional English vocabulary evolves; stay updated on best practices within your team and wider communities. Don’t be afraid to ask for clarification – it’s always better to seek understanding than to risk miscommunication.

redis-cli FT.CREATE my_vector_field --capacity 100000 --orders metric cosine