spencer

spencer is a simple yet get-it-done library that embed any knowledge (e.g., text files) to a searchable database, and allow q&a on the embedded knowledge.

spencer uses redis as vector database, allowing ultra-fast vector similarity searches and retrieval operations.

redis

• https://hub.docker.com/r/redis/redis-stack-server/
• 1st time
  • docker pull redis/redis-stack-server
  • docker run -d --name redis-stack -p 6379:6379 -v ./redis-data:/data redis/redis-stack-server:latest
  • docker stop IMAGE_ID
• onwards
  • docker run -it -p 6379:6379 -v ./redis-data:/data IMAGE_ID
  • docker stop IMAGE_ID

openai

• export OPENAI_API_KEY=KEY

install

• naviage to this folder
• pip install .

usage

embedding

prepare a folder that contains all the knowledge. this folder can be nested.

from spencer import Embedder
r = redis.Redis(host='localhost', port=6379, decode_responses=True)
o = OpenAI(api_key=os.environ.get("OPENAI_API_KEY"))
embedder = Embedder(
    r, o, "text-embedding-3-small", knowledge_dir, max_tokens=2000
)
success = embedder()
print(success)

verify whether the knowledge has been embedded or not via redis-cli

the knowledge keys are in format knowledge:doc_id:chunk_id.

spencer

spencer automatically embed the question, and find the closest k knowledge vector and its text.

spencer uses these k texts as context. further, spencer allows adhoc_context, meaning you can

pass context other than the knolwedge we can just embedded.

r = redis.Redis(host='locahost', port=6379, decode_responses=True)
o = OpenAI(api_key=os.environ.get("OPENAI_API_KEY"))
spencer_client = Spencer(
    r,
    o,
    system_instruction,
    chat_model,
    embedding_model,
    max_context_len,
    max_tokens,
)

question_1 = '...'
resp = spencer_client.answer(question_1, adhoc_context="...")

question_2 = '...'
resp = spencer_client.answer(question_2, adhoc_context="...")

reference

• based on https://platform.openai.com/docs/tutorials/web-qa-embeddings
• based on https://redis.io/docs/get-started/vector-database/