adding hugging face text embedings

embeddings/tei/doc.go

@@ -0,0 +1,8 @@
+/*
+Huggingface Text Embeddings Inference
+https://github.com/huggingface/text-embeddings-inference
+
+package is a wrapper for the Huggingface text embeddings inference project
+that can be run locally for creating vector embeddings.
+*/
+package tei

embeddings/tei/options.go

@@ -0,0 +1,97 @@
+package tei
+
+import (
+	"errors"
+	"runtime"
+	"time"
+
+	client "github.com/gage-technologies/tei-go"
+)
+
+const (
+	_defaultBatchSize       = 512
+	_defaultStripNewLines   = true
+	_defaultTimeNanoSeconds = 60 * 1000000000
+)
+
+var ErrMissingAPIBaseURL = errors.New("missing the API Base URL") //nolint:lll
+
+type Option func(emb *TextEmbeddingsInference)
+
+// WithStripNewLines is an option for specifying the should it strip new lines.
+func WithStripNewLines(stripNewLines bool) Option {
+	return func(p *TextEmbeddingsInference) {
+		p.StripNewLines = stripNewLines
+	}
+}
+
+// WithPoolSize is an option for specifying the number of goroutines.
+func WithPoolSize(poolSize int) Option {
+	return func(p *TextEmbeddingsInference) {
+		p.poolSize = poolSize
+	}
+}
+
+// WithBatchSize is an option for specifying the batch size.
+func WithBatchSize(batchSize int) Option {
+	return func(p *TextEmbeddingsInference) {
+		p.BatchSize = batchSize
+	}
+}
+
+// WithAPIBaseURL adds base url for api.
+func WithAPIBaseURL(url string) Option {
+	return func(emb *TextEmbeddingsInference) {
+		emb.baseURL = url
+	}
+}
+
+// WithHeaders add request headers.
+func WithHeaders(headers map[string]string) Option {
+	return func(emb *TextEmbeddingsInference) {
+		if emb.headers == nil {
+			emb.headers = make(map[string]string, len(headers))
+		}
+		for k, v := range headers {
+			emb.headers[k] = v
+		}
+	}
+}
+
+// WithCookies add request cookies.
+func WithCookies(cookies map[string]string) Option {
+	return func(emb *TextEmbeddingsInference) {
+		if emb.cookies == nil {
+			emb.cookies = make(map[string]string, len(cookies))
+		}
+		for k, v := range cookies {
+			emb.cookies[k] = v
+		}
+	}
+}
+
+// WithTimeout set the request timeout.
+func WithTimeout(dur time.Duration) Option {
+	return func(emb *TextEmbeddingsInference) {
+		emb.timeout = dur
+	}
+}
+
+func applyClientOptions(opts ...Option) (TextEmbeddingsInference, error) {
+	emb := TextEmbeddingsInference{
+		StripNewLines: _defaultStripNewLines,
+		BatchSize:     _defaultBatchSize,
+		timeout:       time.Duration(_defaultTimeNanoSeconds),
+		poolSize:      runtime.GOMAXPROCS(0),
+	}
+	for _, opt := range opts {
+		opt(&emb)
+	}
+	if emb.baseURL == "" {
+		return emb, ErrMissingAPIBaseURL
+	}
+	if emb.client == nil {
+		emb.client = client.NewClient(emb.baseURL, emb.headers, emb.cookies, emb.timeout)
+	}
+	return emb, nil
+}

embeddings/tei/text_embeddings_inference..go

@@ -0,0 +1,84 @@
+package tei
+
+import (
+	"context"
+	"strings"
+	"time"
+
+	client "github.com/gage-technologies/tei-go"
+	"github.com/sourcegraph/conc/pool"
+	"github.com/tmc/langchaingo/embeddings"
+)
+
+type TextEmbeddingsInference struct {
+	client        *client.Client
+	StripNewLines bool
+	BatchSize     int
+	baseURL       string
+	headers       map[string]string
+	cookies       map[string]string
+	timeout       time.Duration
+	poolSize      int
+}
+
+var _ embeddings.Embedder = TextEmbeddingsInference{}
+
+func New(opts ...Option) (TextEmbeddingsInference, error) {
+	emb, err := applyClientOptions(opts...)
+	if err != nil {
+		return emb, err
+	}
+	emb.client = client.NewClient(emb.baseURL, emb.headers, emb.cookies, emb.timeout)
+
+	return emb, nil
+}
+
+// EmbedDocuments creates one vector embedding for each of the texts.
+func (e TextEmbeddingsInference) EmbedDocuments(_ context.Context, texts []string) ([][]float32, error) {
+	batchedTexts := embeddings.BatchTexts(
+		embeddings.MaybeRemoveNewLines(texts, e.StripNewLines),
+		e.BatchSize,
+	)
+
+	emb := make([][]float32, 0, len(texts))
+
+	p := pool.New().WithMaxGoroutines(e.poolSize).WithErrors()
+
+	for _, txt := range batchedTexts {
+		p.Go(func() error {
+			curTextEmbeddings, err := e.client.Embed(strings.Join(txt, " "), false)
+			if err != nil {
+				return err
+			}
+
+			textLengths := make([]int, 0, len(txt))
+			for _, text := range txt {
+				textLengths = append(textLengths, len(text))
+			}
+
+			combined, err := embeddings.CombineVectors(curTextEmbeddings, textLengths)
+			if err != nil {
+				return err
+			}
+
+			emb = append(emb, combined)
+
+			return nil
+		})
+	}
+	return emb, p.Wait()
+}
+
+// EmbedQuery embeds a single text.
+func (e TextEmbeddingsInference) EmbedQuery(_ context.Context, text string) ([]float32, error) {
+	if e.StripNewLines {
+		text = strings.ReplaceAll(text, "\n", " ")
+	}
+
+	emb, err := e.client.Embed(text, false)
+	if err != nil {
+		return nil, err
+	}
+
+	return emb[0], nil
+}