mirror of https://github.com/hwchase17/langchain
mongo parent document retrieval (#12887)
Harrison Chase
7 months ago
committed by
GitHub
parent
e43b4079c8
commit
60d025b83b
@ -0,0 +1,21 @@
|
||||
MIT License
|
||||
|
||||
Copyright (c) 2023 LangChain, Inc.
|
||||
|
||||
Permission is hereby granted, free of charge, to any person obtaining a copy
|
||||
of this software and associated documentation files (the "Software"), to deal
|
||||
in the Software without restriction, including without limitation the rights
|
||||
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
||||
copies of the Software, and to permit persons to whom the Software is
|
||||
furnished to do so, subject to the following conditions:
|
||||
|
||||
The above copyright notice and this permission notice shall be included in all
|
||||
copies or substantial portions of the Software.
|
||||
|
||||
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
||||
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
||||
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
||||
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
||||
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
||||
SOFTWARE.
|
||||
@ -0,0 +1,178 @@
|
||||
# mongo-parent-document-retrieval
|
||||
|
||||
This template performs RAG using MongoDB and OpenAI.
|
||||
It does a more advanced form of RAG called Parent-Document Retrieval.
|
||||
|
||||
In this form of retrieval, a large document is first split into medium sized chunks.
|
||||
From there, those medium size chunks are split into small chunks.
|
||||
Embeddings are created for the small chunks.
|
||||
When a query comes in, an embedding is created for that query and compared to the small chunks.
|
||||
But rather than passing the small chunks directly to the LLM for generation, the medium-sized chunks
|
||||
from whence the smaller chunks came are passed.
|
||||
This helps enable finer-grained search, but then passing of larger context (which can be useful during generation).
|
||||
|
||||
## Environment Setup
|
||||
|
||||
You should export two environment variables, one being your MongoDB URI, the other being your OpenAI API KEY.
|
||||
If you do not have a MongoDB URI, see the `Setup Mongo` section at the bottom for instructions on how to do so.
|
||||
|
||||
```shell
|
||||
export MONGO_URI=...
|
||||
export OPENAI_API_KEY=...
|
||||
```
|
||||
|
||||
## Usage
|
||||
|
||||
To use this package, you should first have the LangChain CLI installed:
|
||||
|
||||
```shell
|
||||
pip install -U langchain-cli
|
||||
```
|
||||
|
||||
To create a new LangChain project and install this as the only package, you can do:
|
||||
|
||||
```shell
|
||||
langchain app new my-app --package mongo-parent-document-retrieval
|
||||
```
|
||||
|
||||
If you want to add this to an existing project, you can just run:
|
||||
|
||||
```shell
|
||||
langchain app add mongo-parent-document-retrieval
|
||||
```
|
||||
|
||||
And add the following code to your `server.py` file:
|
||||
```python
|
||||
from mongo_parent_document_retrieval import chain as mongo_parent_document_retrieval_chain
|
||||
|
||||
add_routes(app, mongo_parent_document_retrieval_chain, path="/mongo-parent-document-retrieval")
|
||||
```
|
||||
|
||||
(Optional) Let's now configure LangSmith.
|
||||
LangSmith will help us trace, monitor and debug LangChain applications.
|
||||
LangSmith is currently in private beta, you can sign up [here](https://smith.langchain.com/).
|
||||
If you don't have access, you can skip this section
|
||||
|
||||
|
||||
```shell
|
||||
export LANGCHAIN_TRACING_V2=true
|
||||
export LANGCHAIN_API_KEY=<your-api-key>
|
||||
export LANGCHAIN_PROJECT=<your-project> # if not specified, defaults to "default"
|
||||
```
|
||||
|
||||
If you DO NOT already have a Mongo Search Index you want to connect to, see `MongoDB Setup` section below before proceeding.
|
||||
Note that because Parent Document Retrieval uses a different indexing strategy, it's likely you will want to run this new setup.
|
||||
|
||||
If you DO have a MongoDB Search index you want to connect to, edit the connection details in `mongo_parent_document_retrieval/chain.py`
|
||||
|
||||
If you are inside this directory, then you can spin up a LangServe instance directly by:
|
||||
|
||||
```shell
|
||||
langchain serve
|
||||
```
|
||||
|
||||
This will start the FastAPI app with a server is running locally at
|
||||
[http://localhost:8000](http://localhost:8000)
|
||||
|
||||
We can see all templates at [http://127.0.0.1:8000/docs](http://127.0.0.1:8000/docs)
|
||||
We can access the playground at [http://127.0.0.1:8000/mongo-parent-document-retrieval/playground](http://127.0.0.1:8000/mongo-parent-document-retrieval/playground)
|
||||
|
||||
We can access the template from code with:
|
||||
|
||||
```python
|
||||
from langserve.client import RemoteRunnable
|
||||
|
||||
runnable = RemoteRunnable("http://localhost:8000/mongo-parent-document-retrieval")
|
||||
```
|
||||
|
||||
For additional context, please refer to [this notebook](https://colab.research.google.com/drive/1cr2HBAHyBmwKUerJq2if0JaNhy-hIq7I#scrollTo=TZp7_CBfxTOB).
|
||||
|
||||
|
||||
## MongoDB Setup
|
||||
|
||||
Use this step if you need to setup your MongoDB account and ingest data.
|
||||
We will first follow the standard MongoDB Atlas setup instructions [here](https://www.mongodb.com/docs/atlas/getting-started/).
|
||||
|
||||
1. Create an account (if not already done)
|
||||
2. Create a new project (if not already done)
|
||||
3. Locate your MongoDB URI.
|
||||
|
||||
This can be done by going to the deployement overview page and connecting to you database
|
||||
|
||||
|
||||
|
||||
We then look at the drivers available
|
||||
|
||||
|
||||
|
||||
Among which we will see our URI listed
|
||||
|
||||
|
||||
|
||||
Let's then set that as an environment variable locally:
|
||||
|
||||
```shell
|
||||
export MONGO_URI=...
|
||||
```
|
||||
|
||||
4. Let's also set an environment variable for OpenAI (which we will use as an LLM)
|
||||
|
||||
```shell
|
||||
export OPENAI_API_KEY=...
|
||||
```
|
||||
|
||||
5. Let's now ingest some data! We can do that by moving into this directory and running the code in `ingest.py`, eg:
|
||||
|
||||
```shell
|
||||
python ingest.py
|
||||
```
|
||||
|
||||
Note that you can (and should!) change this to ingest data of your choice
|
||||
|
||||
6. We now need to set up a vector index on our data.
|
||||
|
||||
We can first connect to the cluster where our database lives
|
||||
|
||||
|
||||
|
||||
We can then navigate to where all our collections are listed
|
||||
|
||||
|
||||
|
||||
We can then find the collection we want and look at the search indexes for that collection
|
||||
|
||||
|
||||
|
||||
That should likely be empty, and we want to create a new one:
|
||||
|
||||
|
||||
|
||||
We will use the JSON editor to create it
|
||||
|
||||
|
||||
|
||||
And we will paste the following JSON in:
|
||||
|
||||
```text
|
||||
{
|
||||
"mappings": {
|
||||
"dynamic": true,
|
||||
"fields": {
|
||||
"doc_level": [
|
||||
{
|
||||
"type": "token"
|
||||
}
|
||||
],
|
||||
"embedding": {
|
||||
"dimensions": 1536,
|
||||
"similarity": "cosine",
|
||||
"type": "knnVector"
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
|
||||
From there, hit "Next" and then "Create Search Index". It will take a little bit but you should then have an index over your data!
|
||||
|
||||
@ -0,0 +1,59 @@
|
||||
import os
|
||||
import uuid
|
||||
|
||||
from langchain.document_loaders import PyPDFLoader
|
||||
from langchain.embeddings import OpenAIEmbeddings
|
||||
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
||||
from langchain.vectorstores import MongoDBAtlasVectorSearch
|
||||
from pymongo import MongoClient
|
||||
|
||||
PARENT_DOC_ID_KEY = "parent_doc_id"
|
||||
|
||||
|
||||
def parent_child_splitter(data, id_key=PARENT_DOC_ID_KEY):
|
||||
parent_splitter = RecursiveCharacterTextSplitter(chunk_size=2000)
|
||||
# This text splitter is used to create the child documents
|
||||
# It should create documents smaller than the parent
|
||||
child_splitter = RecursiveCharacterTextSplitter(chunk_size=400)
|
||||
documents = parent_splitter.split_documents(data)
|
||||
doc_ids = [str(uuid.uuid4()) for _ in documents]
|
||||
|
||||
docs = []
|
||||
for i, doc in enumerate(documents):
|
||||
_id = doc_ids[i]
|
||||
sub_docs = child_splitter.split_documents([doc])
|
||||
for _doc in sub_docs:
|
||||
_doc.metadata[id_key] = _id
|
||||
_doc.metadata["doc_level"] = "child"
|
||||
docs.extend(sub_docs)
|
||||
doc.metadata[id_key] = _id
|
||||
doc.metadata["doc_level"] = "parent"
|
||||
return documents, docs
|
||||
|
||||
|
||||
MONGO_URI = os.environ["MONGO_URI"]
|
||||
|
||||
# Note that if you change this, you also need to change it in `rag_mongo/chain.py`
|
||||
DB_NAME = "langchain-test-2"
|
||||
COLLECTION_NAME = "test"
|
||||
ATLAS_VECTOR_SEARCH_INDEX_NAME = "default"
|
||||
EMBEDDING_FIELD_NAME = "embedding"
|
||||
client = MongoClient(MONGO_URI)
|
||||
db = client[DB_NAME]
|
||||
MONGODB_COLLECTION = db[COLLECTION_NAME]
|
||||
|
||||
if __name__ == "__main__":
|
||||
# Load docs
|
||||
loader = PyPDFLoader("https://arxiv.org/pdf/2303.08774.pdf")
|
||||
data = loader.load()
|
||||
|
||||
# Split docs
|
||||
parent_docs, child_docs = parent_child_splitter(data)
|
||||
|
||||
# Insert the documents in MongoDB Atlas Vector Search
|
||||
_ = MongoDBAtlasVectorSearch.from_documents(
|
||||
documents=parent_docs + child_docs,
|
||||
embedding=OpenAIEmbeddings(disallowed_special=()),
|
||||
collection=MONGODB_COLLECTION,
|
||||
index_name=ATLAS_VECTOR_SEARCH_INDEX_NAME,
|
||||
)
|
||||
@ -0,0 +1,3 @@
|
||||
from mongo_parent_document_retrieval.chain import chain
|
||||
|
||||
__all__ = ["chain"]
|
||||
@ -0,0 +1,91 @@
|
||||
import os
|
||||
|
||||
from langchain.chat_models import ChatOpenAI
|
||||
from langchain.embeddings import OpenAIEmbeddings
|
||||
from langchain.prompts import ChatPromptTemplate
|
||||
from langchain.pydantic_v1 import BaseModel
|
||||
from langchain.schema.document import Document
|
||||
from langchain.schema.output_parser import StrOutputParser
|
||||
from langchain.schema.runnable import RunnableParallel, RunnablePassthrough
|
||||
from langchain.vectorstores import MongoDBAtlasVectorSearch
|
||||
from pymongo import MongoClient
|
||||
|
||||
MONGO_URI = os.environ["MONGO_URI"]
|
||||
PARENT_DOC_ID_KEY = "parent_doc_id"
|
||||
# Note that if you change this, you also need to change it in `rag_mongo/chain.py`
|
||||
DB_NAME = "langchain-test-2"
|
||||
COLLECTION_NAME = "test"
|
||||
ATLAS_VECTOR_SEARCH_INDEX_NAME = "default"
|
||||
EMBEDDING_FIELD_NAME = "embedding"
|
||||
client = MongoClient(MONGO_URI)
|
||||
db = client[DB_NAME]
|
||||
MONGODB_COLLECTION = db[COLLECTION_NAME]
|
||||
|
||||
|
||||
vector_search = MongoDBAtlasVectorSearch.from_connection_string(
|
||||
MONGO_URI,
|
||||
DB_NAME + "." + COLLECTION_NAME,
|
||||
OpenAIEmbeddings(disallowed_special=()),
|
||||
index_name=ATLAS_VECTOR_SEARCH_INDEX_NAME,
|
||||
)
|
||||
|
||||
|
||||
def retrieve(query: str):
|
||||
results = vector_search.similarity_search(
|
||||
query,
|
||||
k=4,
|
||||
pre_filter={"doc_level": {"$eq": "child"}},
|
||||
post_filter_pipeline=[
|
||||
{"$project": {"embedding": 0}},
|
||||
{
|
||||
"$lookup": {
|
||||
"from": COLLECTION_NAME,
|
||||
"localField": PARENT_DOC_ID_KEY,
|
||||
"foreignField": PARENT_DOC_ID_KEY,
|
||||
"as": "parent_context",
|
||||
"pipeline": [
|
||||
{"$match": {"doc_level": "parent"}},
|
||||
{"$limit": 1},
|
||||
{"$project": {"embedding": 0}},
|
||||
],
|
||||
}
|
||||
},
|
||||
],
|
||||
)
|
||||
parent_docs = []
|
||||
parent_doc_ids = set()
|
||||
for result in results:
|
||||
res = result.metadata["parent_context"][0]
|
||||
text = res.pop("text")
|
||||
# This causes serialization issues.
|
||||
res.pop("_id")
|
||||
parent_doc = Document(page_content=text, metadata=res)
|
||||
if parent_doc.metadata[PARENT_DOC_ID_KEY] not in parent_doc_ids:
|
||||
parent_doc_ids.add(parent_doc.metadata[PARENT_DOC_ID_KEY])
|
||||
parent_docs.append(parent_doc)
|
||||
return parent_docs
|
||||
|
||||
|
||||
# RAG prompt
|
||||
template = """Answer the question based only on the following context:
|
||||
{context}
|
||||
Question: {question}
|
||||
"""
|
||||
prompt = ChatPromptTemplate.from_template(template)
|
||||
|
||||
# RAG
|
||||
model = ChatOpenAI()
|
||||
chain = (
|
||||
RunnableParallel({"context": retrieve, "question": RunnablePassthrough()})
|
||||
| prompt
|
||||
| model
|
||||
| StrOutputParser()
|
||||
)
|
||||
|
||||
|
||||
# Add typing for input
|
||||
class Question(BaseModel):
|
||||
__root__: str
|
||||
|
||||
|
||||
chain = chain.with_types(input_type=Question)
|
||||
@ -0,0 +1,27 @@
|
||||
[tool.poetry]
|
||||
name = "mongo-parent-document-retrieval"
|
||||
version = "0.0.1"
|
||||
description = ""
|
||||
authors = []
|
||||
readme = "README.md"
|
||||
|
||||
[tool.poetry.dependencies]
|
||||
python = ">=3.8.1,<4.0"
|
||||
langchain = ">=0.0.313, <0.1"
|
||||
openai = "^0.28.1"
|
||||
pymongo = "^4.6.0"
|
||||
pypdf = "^3.17.0"
|
||||
tiktoken = "^0.5.1"
|
||||
|
||||
[tool.poetry.group.dev.dependencies]
|
||||
langchain-cli = ">=0.0.4"
|
||||
fastapi = "^0.104.0"
|
||||
sse-starlette = "^1.6.5"
|
||||
|
||||
[tool.langserve]
|
||||
export_module = "mongo_parent_document_retrieval"
|
||||
export_attr = "chain"
|
||||
|
||||
[build-system]
|
||||
requires = ["poetry-core"]
|
||||
build-backend = "poetry.core.masonry.api"
|
||||
Loading…
Reference in New Issue