langchain-dewey 0.2.0

LangChain integration for Dewey — retriever, vector store, and research tool

langchain-dewey

LangChain integration for Dewey — retriever, vector store, and research tool.

Installation

pip install langchain-dewey

Components

DeweyRetriever

Drop-in LangChain retriever backed by Dewey's hybrid semantic + BM25 search.

from langchain_dewey import DeweyRetriever
from langchain_openai import ChatOpenAI
from langchain.chains import RetrievalQA

retriever = DeweyRetriever(
    api_key="dwy_live_...",
    collection_id="3f7a1b2c-...",
    k=8,
)

qa = RetrievalQA.from_chain_type(llm=ChatOpenAI(), retriever=retriever)
answer = qa.invoke("What are the main findings?")

Each returned Document carries citation metadata:

Field	Description
score	Relevance score (0–1)
document_id	Dewey document ID
filename	Original filename
section_id	Section ID
section_title	Section heading
section_level	Heading depth (1 = top-level)

DeweyVectorStore

Full VectorStore implementation. Dewey manages its own embeddings, so the embedding argument is accepted for interface compatibility but unused.

Wrap an existing collection:

from langchain_dewey import DeweyVectorStore

store = DeweyVectorStore(
    api_key="dwy_live_...",
    collection_id="3f7a1b2c-...",
)
docs = store.similarity_search("retrieval augmented generation", k=5)

Build from texts:

store = DeweyVectorStore.from_texts(
    texts=["Neural networks learn via backpropagation.", "..."],
    embedding=None,
    api_key="dwy_live_...",
    collection_name="my-docs",
)
docs = store.similarity_search("how does training work?")

add_texts uploads each string as a .txt document and blocks until Dewey finishes processing them (configurable via poll_interval and poll_timeout).

With LangChain document loaders:

from langchain_community.document_loaders import PyPDFLoader

loader = PyPDFLoader("research_paper.pdf")
pages = loader.load_and_split()

store = DeweyVectorStore.from_documents(
    pages,
    embedding=None,
    api_key="dwy_live_...",
    collection_name="research",
)

create_research_tool

Creates a LangChain @tool that runs a full Dewey research query — searching, reading, and synthesising across multiple documents — for use with agents.

from langchain_dewey import create_research_tool
from langchain_openai import ChatOpenAI
from langchain.agents import AgentExecutor, create_openai_tools_agent
from langchain_core.prompts import ChatPromptTemplate

research = create_research_tool(
    api_key="dwy_live_...",
    collection_id="3f7a1b2c-...",
    depth="balanced",  # "quick" | "balanced" | "deep" | "exhaustive"
)

llm = ChatOpenAI(model="gpt-4o")
prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful research assistant."),
    ("human", "{input}"),
    ("placeholder", "{agent_scratchpad}"),
])
agent = create_openai_tools_agent(llm, [research], prompt)
executor = AgentExecutor(agent=agent, tools=[research])
executor.invoke({"input": "Summarise the key findings across all documents."})

Requirements

• Python 3.9+
• meetdewey >= 1.0
• langchain-core >= 0.3

Development

pip install -e ".[dev]"
pytest