Feature Tutorials

Embeddings & Vector Search

Generate dense vector embeddings for semantic search, document clustering, and RAG pipelines.

Overview

Vector embeddings convert text into numerical arrays (vectors) that capture semantic meaning. Texts with similar meanings produce vectors that are close together in multi-dimensional space.

Ollama provides high-speed local embedding models such as nomic-embed-text and all-minilm.

1. Generating Embeddings

Use embed with embedRequest:

{-# LANGUAGE OverloadedStrings #-}
module Main where

import Ollama

main :: IO ()
main = do
  client <- defaultClient

  -- Generate embedding for a single text
  let req = embedRequest "nomic-embed-text" ["Haskell is a purely functional programming language."]
  res <- embed client req

  case res of
    Left err   -> print err
    Right resp -> do
      let vectors = erEmbeddings resp
      putStrLn $ "Generated " <> show (length vectors) <> " vector(s)"
      case vectors of
        (vec : _) -> do
          putStrLn $ "Vector dimensions: " <> show (length vec)
          putStrLn $ "First 5 dimensions: " <> show (take 5 vec)
        [] -> putStrLn "No embeddings returned"

2. Batch Embedding & Semantic Similarity Search

Generate embeddings for multiple documents in a single request and rank them using cosine similarity:

{-# LANGUAGE OverloadedStrings #-}
module Main where

import Data.List (sortBy)
import Data.Ord (Down (..))
import Data.Text (Text)
import Ollama

-- Calculate cosine similarity between two vectors
cosineSimilarity :: [Double] -> [Double] -> Double
cosineSimilarity u v =
  let dotProduct = sum $ zipWith (*) u v
      normA = sqrt . sum $ map (^ (2 :: Int)) u
      normB = sqrt . sum $ map (^ (2 :: Int)) v
   in if normA == 0 || normB == 0 then 0 else dotProduct / (normA * normB)

documents :: [Text]
documents =
  [ "GHC compiles Haskell source code to native machine instructions."
  , "Bananas and apples are nutritious fruits."
  , "Type inference with Hindley-Milner algorithms guarantees static safety."
  , "The weather forecast calls for rain tomorrow."
  ]

main :: IO ()
main = do
  client <- defaultClient

  -- 1. Embed documents
  docRes <- embed client $ embedRequest "nomic-embed-text" documents

  -- 2. Embed user search query
  queryRes <- embed client $ embedRequest "nomic-embed-text" ["How does Haskell typing work?"]

  case (docRes, queryRes) of
    (Right dResp, Right qResp) -> do
      case (erEmbeddings dResp, erEmbeddings qResp) of
        (docVecs, queryVec : _) -> do
          let scored = zip documents (map (cosineSimilarity queryVec) docVecs)
              ranked = sortBy (\(_, s1) (_, s2) -> compare (Down s1) (Down s2)) scored

          putStrLn "--- Top Search Results ---"
          mapM_ (\(doc, score) -> putStrLn $ "[" <> show (round (score * 100) :: Int) <> "% match] " <> show doc) ranked
        _ -> putStrLn "Missing embeddings"
    _ -> putStrLn "Error computing embeddings"