This project creates an api to query with GPT on your files, TODO: pdf integration This project also provides a text parser, for parsing into chunks to be embbeded. It also includes an API to interact with OpenAI's GPT-3 and text embeddings services.
- Suport for text-davinci-003
- Query you files using embedings and answer your questions using gpt
- Connect to OpenAI API for embeddings and completions
- Store paragraph embeddings in a PostgreSQL database using the pgvector extension
- Find most similar paragraphs based on embeddings
- Golang package structure with internal and pkg directories
- Put your OPEN IA KEY in the docker-compose.yml
- In main.go put the file you want to use
- At the moment we are usinig text-davinci-003 that only accept 2048 tokens, so we neet to tweek the config to match that pre requists
docker-compose build
docker-compose up -dYou can make an API call to get completions using a REST client or a tool like curl. Here's an example API call to a locally running server:
- Endpoint http://localhost:3000/api/completition
- body: {"query": "string"}
curl -X POST http://localhost:3000/api/completition \
-H "Content-Type: application/json" \
-d '{"query": "what is the article about?"}'ParseTxtInChunks function reads a text file located at the given filePath and splits its content into chunks of chunkSize words with an overlap number of words between each chunk. It returns a slice of strings, where each string represents a chunk of text.
The SaveParagraph method first checks if the content already exists in the database by calling the ContentExists method from the repo. If the content exists, it returns immediately, not saving the paragraph again.
If the content does not exist, the method retrieves the paragraph's embedding from the OpenAI API using the GetEmbedding method. Then, it creates a new models.Paragraph instance with the content and its embedding, and saves the paragraph in the repository using the SaveParagraph method.
func (a *application) SaveParagraph(ctx context.Context, content string) error {
// Check if the content already exists in the database
exists, err := a.repo.ContentExists(ctx, content)
if err != nil {
return err
}
// If the content doesn't exist, save the paragraph with its embedding
if !exists {
embedding, err := a.openApi.GetEmbedding(content)
if err != nil {
return err
}
paragraph := models.NewParagraph(content, embedding)
return a.repo.SaveParagraph(ctx, ¶graph)
}
return nil
}The GetGptResposeWithContext method first retrieves the embedding for the given question using the OpenAI API. It then finds the most similar paragraphs to the question by calling the GetMostSimilarVectors method from the repo.
It builds a context string by concatenating the content of the most similar paragraphs, limiting the total tokens to 1500. The context string and the question are then formatted into a GPT-3 prompt. Finally, it calls the OpenAI API to get a completion based on the generated prompt, using the GetCompletion method.
func (a *application) GetGptResposeWithContext(ctx context.Context, question string, model string) (string, error) {
// Get the embedding for the question
embedding, err := a.openApi.GetEmbedding(question)
if err != nil {
return "", err
}
// Find the most similar paragraphs
results, err := a.repo.GetMostSimilarVectors(ctx, embedding, 5)
if err != nil {
return "", err
}
// Build the context string
context := ""
tokens := 0
for _, result := range results {
if tokens >= 2000 {
break
}
context = context + result.Content + "\n"
tokens = tokens + result.TokenCount
}
// Format the GPT-3 prompt
prompt := fmt.Sprintf(`
// ...
Context sextions: %s,
Question: %s
`, context, question)
// Get the completion from the OpenAI API
completion, err := a.openApi.GetCompletion(prompt, 1500, 0.5, model)
if err != nil {
return "", err
}
return completion, nil
}Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.
Please make sure to update tests as appropriate.