![Logo](https://github.com/DewyKB/dewy/raw/main/images/logo.png?raw=true)
Opinionated knowledge extraction and semantic retrieval for Gen AI applications.
Explore the docs »
Report Bug
·
Request Feature
Dewy helps you build AI agents and RAG applications by managing the extraction of knowledge from your documents and implementing semantic search over the extracted content. Load your documents and Dewy takes care of parsing, chunking, summarizing, and indexing for retrieval. Dewy builds on the lessons of putting real Gen AI applications into production so you can focus on getting 💩 done, rather than comparing vector databases and building data extraction infrastructure.
Below is the typical architecture of an AI agent performing RAG. Dewy handles all of the parts shown in brown so you can focus on your application -- the parts in green.
To get a local copy up and running follow these steps.
-
(Optional) Start a
pgvector
instance to persist your dataDewy uses a vector database to store metadata about the documents you've loaded as well as embeddings used to provide semantic search results.
docker run -d \ -p 5432:5432 \ -e POSTGRES_DB=dewydb \ -e POSTGRES_USER=dewydbuser \ -e POSTGRES_PASSWORD=dewydbpwd \ -e POSTGRES_HOST_AUTH_METHOD=trust \ ankane/pgvector
If you already have an instance of
pgvector
you can create a database for Dewy and configure Dewy use it using theDB
env var (see below). -
Install Dewy
pip install dewy
This will install Dewy in your local Python environment.
-
Configure Dewy. Dewy will read env vars from an
.env
file if provided. You can also set these directly in the environment, for example when configuring an instance running in docker / kubernetes.# ~/.env ENVIRONMENT=LOCAL DB=postgresql://... OPENAI_API_KEY=...
-
Fire up Dewy
dewy
Dewy includes an admin console you can use to create collections, load documents, and run test queries.
open http://localhost:8000/admin
-
Install the API client library
npm install dewy-ts
-
Connect to an instance of Dewy
import { Dewy } from 'dewy_ts'; const dewy = new Dewy()
-
Add documents
await dewy.kb.addDocument({ collection_id: 1, url: “https://arxiv.org/abs/2005.11401”, })
-
Retrieve document chunks for LLM prompting
const context = await dewy.kb.retrieveChunks({ collection_id: 1, query: "tell me about RAG", n: 10, }); // Minimal prompt example const prompt = [ { role: 'system', content: `You are a helpful assistant. You will take into account any CONTEXT BLOCK that is provided in a conversation. START CONTEXT BLOCK ${context.results.map((c: any) => c.chunk.text).join("\n")} END OF CONTEXT BLOCK `, }, ] // Using OpenAI to generate responses const response = await openai.chat.completions.create({ model: 'gpt-3.5-turbo', stream: true, messages: [...prompt, [{role: 'user': content: 'Tell me about RAG'}]] })
-
Install the API client library
pip install dewy-client
-
Connect to an instance of Dewy
from dewy_client import Client dewy = Client(base_url="http://localhost:8000")
-
Add documents
from dewy_client.api.kb import add_document from dewy_client.models import AddDocumentRequest await add_document.asyncio(client=dewy, body=AddDocumentRequest( collection_id = 1, url = “https://arxiv.org/abs/2005.11401”, ))
-
Retrieve document chunks for LLM prompting
from dewy_client.api.kb import retrieve_chunks from dewy_client.modles import RetrieveRequest chunks = await retrieve_chunks.asyncio(client=dewy, body=RetrieveRequest( collection_id = 1, query = "tell me about RAG", n = 10, )) # Minimal prompt example prompt = f""" You will take into account any CONTEXT BLOCK that is provided in a conversation. START CONTEXT BLOCK {"\n".join([chunk.text for chunk in chunks.text_chunks])} END OF CONTEXT BLOCK """
See [python-langchain.ipynb'](demos/python-langchain-notebook/python-langchain.ipynb) for an example using Dewy in LangChain, including an implementation of LangChain's
BaseRetriever` backed by Dewy.
Dewy is under active development. This is an overview of our current roadmap - please 👍 issues that are important to you. Don't see a feature that would make Dewy better for your application - create a feature request!
- Support more document formats (ie Markdown, DOCX, HTML)
- Support more types of chunk extractors
- Multi-modal search over images, tables, audio, etc.
- Integrations with LangChain, LlamaIndex, Haystack, etc.
- Support flexible result ranking (ie rag-fusion, mmr, etc).
- Provide metrics around which chunks are used, relevance scores, etc.
- Query history and explorer in the UI.
- Multi-tenancy
- Hybrid search
Contributions are what make the open source community such an amazing place to learn, inspire, and create. Any contributions you make are greatly appreciated.
If you have a suggestion that would make this better, please fork the repo and create a pull request. You can also simply open an issue with the tag "enhancement". Don't forget to give the project a star! Thanks again!
- Fork the Project
- Create your Feature Branch (
git checkout -b feature/AmazingFeature
) - Commit your Changes (
git commit -m 'Add some AmazingFeature'
) - Push to the Branch (
git push origin feature/AmazingFeature
) - Open a Pull Request
- Clone the repo
git clone https://github.com/DewyKB/dewy.git
- Install Python packages
poetry install
- Configure Dewy.
Dewy will read env vars from an
.env
file if provided. You can also set these directly in the environment, for example when configuring an instance running in docker / kubernetes.cat > .env << EOF ENVIRONMENT=LOCAL DB=postgresql://... OPENAI_API_KEY=... EOF
- Build the frontend
cd frontend && npm install && npm run build
- Build the client
cd dewy-client && poetry install
- Run the Dewy service
poetry run dewy
Some skeleton code based on best practices from https://github.com/zhanymkanov/fastapi-best-practices.
The following commands run tests and apply linting.
If you're in a poetry shell
, you can omit the poetry run
:
- Running tests:
poetry run pytest
- Linting (and formatting):
poetry run ruff check --fix
- Formatting:
poetry run ruff format
- Type Checking:
poetry run mypy dewy
To regenerate the OpenAPI spec and client libraries:
poetry poe extract-openapi
poetry poe update-client
- Look at the draft release to determine the suggested next version.
- Create a PR updating the following locations to that version:
a.
pyproject.toml
fordewy
b.dewy-client/pyproject.toml
fordewy-client
c. API version indewy/config.py
d.openapi.yaml
anddewy-client
by runningpoe extract-openapi
andpoe update-client
. - Once that PR is in, edit the draft release, make sure the version and tag match what you selected in step 1 (and used in the PR), check "Set as a pre-release" (will be updated by the release automation) and choose to publish the release.
- The release automation should kick in and work through the release steps. It will need approval for the pypi deployment environment to publish the
dewy
anddewy-client
packages.
Distributed under the Apache 2 License. See LICENSE.txt
for more information.