AwaDB - AI Native Database for embedding vectors

Easily Use - No boring database schema definition. No need to pay attention to vector indexing details.

Realtime Search - Lock free realtime index keeps new data fresh with millisecond level latency. No wait no manual operation.

Stability - AwaDB builds upon over 5 years experience running production workloads at scale using a system called Vearch, combined with best-of-breed ideas and practices from the community.

Run awadb locally on Mac OSX or Linux

First install awadb:

pip3 install awadb

Then use as below:

import awadb
# 1. Initialize awadb client!
awadb_client = awadb.Client()

# 2. Create table
awadb_client.Create("test_llm1") 

# 3. Add sentences, the sentence is embedded with SentenceTransformer by default
#    You can also embed the sentences all by yourself with OpenAI or other LLMs
awadb_client.Add([{'embedding_text':'The man is happy'}, {'source' : 'pic1'}])
awadb_client.Add([{'embedding_text':'The man is very happy'}, {'source' : 'pic2'}])
awadb_client.Add([{'embedding_text':'The cat is happy'}, {'source' : 'pic3'}])
awadb_client.Add([{'embedding_text':'The man is eating'}, {'source':'pic4'}])

# 4. Search the most Top3 sentences by the specified query
query = "The man is happy"
results = awadb_client.Search(query, 3)

# Output the results
print(results)

Here the text is embedded by SentenceTransformer which is supported by Hugging Face
More detailed python local library usage you can read here

Run AwaDB as a service

If you are on the Windows platform or want a awadb service, you can download and deploy the awadb docker. The installation of awadb docker please see here

Python Usage

First, Install gRPC and awadb service python client as below:

pip3 install grpcio
pip3 install awadb-client

A simple example as below:

# Import the package and module
from awadb_client import Awa

# Initialize awadb client
client = Awa()

# Add dict with vector to table 'example1'
client.add("example1", {'name':'david', 'feature':[1.3, 2.5, 1.9]})
client.add("example1", {'name':'jim', 'feature':[1.1, 1.4, 2.3]})

# Search
results = client.search("example1", [1.0, 2.0, 3.0])

# Output results
print(results)

# '_id' is the primary key of each document
# It can be specified clearly when adding documents
# Here no field '_id' is specified, it is generated by the awadb server 
db_name: "default"
table_name: "example1"
results {
  total: 2
  msg: "Success"
  result_items {
    score: 0.860000074
    fields {
      name: "_id" 
      value: "64ddb69d-6038-4311-9118-605686d758d9"
    }
    fields {
      name: "name"
      value: "jim"
    }
  }
  result_items {
    score: 1.55
    fields {
      name: "_id"
      value: "f9f3035b-faaf-48d4-a947-801416c005b3"
    }
    fields {
      name: "name"
      value: "david"
    }
  }
}
result_code: SUCCESS

More python sdk for service is here

RESTful Usage

# add documents to table 'test' of db 'default', no need to create table first
curl -H "Content-Type: application/json" -X POST -d '{"db":"default", "table":"test", "docs":[{"_id":1, "name":"lj", "age":23, "f":[1,0]},{"_id":2, "name":"david", "age":32, "f":[1,2]}]}' http://localhost:8080/add

# search documents by the vector field 'f' of the value '[1, 1]'
curl -H "Content-Type: application/json" -X POST -d '{"db":"default", "table":"test", "vector_query":{"f":[1, 1]}}' http://localhost:8080/search

More detailed RESTful API is here

What are the Embeddings?

Any unstructured data(image/text/audio/video) can be transferred to vectors which are generally understanded by computers through AI(LLMs or other deep neural networks).

For example, "The man is happy"-this sentence can be transferred to a 384-dimension vector(a list of numbers [0.23, 1.98, ....]) by SentenceTransformer language model. This process is called embedding.

More detailed information about embeddings can be read from OpenAI

Awadb uses Sentence Transformers to embed the sentence by default, while you can also use OpenAI or other LLMs to do the embeddings according to your needs.