Skip to main content

Vectors

A vector (or embedding) is a list of numbers generated by an embedding model to capture the semantic essence of unstructured data — text, an image, audio. The key idea: semantically similar items produce similar vectors, so you can compare meaning instead of matching exact keywords.

“a fast red car”🖼️ photo.jpgEmbeddingmodel[0.91, 0.12, 0.44, …]Raw dataModelVector
An embedding model turns text or images into a vector.

Similar meaning, similar vector​

Two paraphrases of the same sentence, or two photos of the same object, land close together in vector space; unrelated things land far apart. That closeness is exactly what Ahnlich searches on.

“fast red car”“crimson automobile”“bowl of soup”vector space (2-D projection)
Similar meanings land close together; unrelated ones fall far apart.
Text
"a fast red car"              → [0.91, 0.12, 0.44, 0.03]
"a speedy crimson automobile" → [0.89, 0.15, 0.41, 0.05]   ← close ⇒ similar
"a bowl of soup"              → [0.02, 0.77, 0.10, 0.63]    ← far   ⇒ unrelated

How vectors flow through Ahnlich​

Working with vectors is a three-step loop:

  1. Store — turn your data into vectors and insert them into a store. Compute the vectors yourself, or send raw text/images to an AI store and let Ahnlich embed them.
  2. Search — turn your query into a vector with the same model, then run a similarity search.
  3. Retrieve — Ahnlich ranks stored vectors by the chosen similarity metric and returns the nearest, optionally narrowed by metadata predicates.

At scale, an index keeps step 3 fast so you don't compare against every vector.

Dimension​

The dimension is the length of the vector. It's fixed when you create a store, and every key you insert must have exactly that many values — a 4-dimensional store only accepts 4-number keys. The dimension is set by whatever produced the vector (your embedding model, feature extractor, etc.).

Dense vectors​

Ahnlich works with dense vectors: fixed-length lists of real numbers where (nearly) every dimension carries meaning — for example, a 384-dimensional embedding from a language model. Dense vectors are semantically rich but not human-readable; a single number rarely means anything on its own.

Some systems also use sparse vectors — very high-dimensional, mostly-zero vectors where each non-zero slot is a keyword weight (interpretable, but blind to synonyms). Ahnlich focuses on dense vectors.

Match the metric to the model​

Embedding models are trained for a particular notion of distance, so your similarity metric should match how the model was trained — cosine similarity for most text/vision embeddings, Euclidean or dot product where magnitude matters. Using the wrong metric quietly degrades result quality.

Convert each product photo into a vector that captures its visual features — shape, colour, object type. Similar images get nearby vectors, so a new photo surfaces visually similar products, and the same item is recognised across different lighting or angles — all without a single matching filename or tag.