Multi modal RAG template

templates/rag-multi-modal/LICENSE

@@ -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.

templates/rag-multi-modal/README.md

@@ -0,0 +1,77 @@
+# rag-multi-modal
+
+This template performs RAG on multi-modal data, such as a PDF with text and images.
+
+XXX
+
+## Environment Setup
+
+Set the `OPENAI_API_KEY` environment variable to access the OpenAI models.
+
+This uses [Unstructured](https://unstructured-io.github.io/unstructured/) for PDF parsing, which requires some system-level package installations. 
+
+On Mac, you can install the necessary packages with the following:
+
+```shell
+brew install tesseract poppler
+```
+
+## Usage
+
+To use this package, you should first have the LangChain CLI installed:
+
+```shell
+pip install -U "langchain-cli[serve]"
+```
+
+To create a new LangChain project and install this as the only package, you can do:
+
+```shell
+langchain app new my-app --package rag-multi-modal
+```
+
+If you want to add this to an existing project, you can just run:
+
+```shell
+langchain app add rag-multi-modal
+```
+
+And add the following code to your `server.py` file:
+```python
+from rag_semi_structured import chain as rag_semi_structured_chain
+
+add_routes(app, rag_semi_structured_chain, path="/rag-multi-modal")
+```
+
+(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 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/rag-multi-modal/playground](http://127.0.0.1:8000/rag-multi-modal/playground)  
+
+We can access the template from code with:
+
+```python
+from langserve.client import RemoteRunnable
+
+runnable = RemoteRunnable("http://localhost:8000/rag-multi-modal")
+```
+
+For more details on how to connect to the template, refer to the Jupyter notebook `rag-multi-modal`.

templates/rag-multi-modal/pyproject.toml

@@ -0,0 +1,27 @@
+[tool.poetry]
+name = "rag-semi-structured"
+version = "0.1.0"
+description = ""
+authors = [
+    "Lance Martin <lance@langchain.dev>",
+]
+readme = "README.md"
+
+[tool.poetry.dependencies]
+python = ">=3.8.1,<4.0"
+langchain = ">=0.0.325"
+tiktoken = ">=0.5.1"
+chromadb = ">=0.4.14"
+openai = ">=0.27.9"
+unstructured = ">=0.10.19"
+pdf2image = ">=1.16.3"
+
+[tool.langserve]
+export_module = "rag_semi_structured"
+export_attr = "chain"
+
+[build-system]
+requires = [
+    "poetry-core",
+]
+build-backend = "poetry.core.masonry.api"

templates/rag-multi-modal/rag_multi_modal.ipynb

@@ -0,0 +1,51 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "30fc2c27",
+   "metadata": {},
+   "source": [
+    "## Run Template\n",
+    "\n",
+    "In `server.py`, set -\n",
+    "```\n",
+    "add_routes(app, chain_rag_conv, path=\"/multi-modal-rag\")\n",
+    "```"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "65f5b560",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langserve.client import RemoteRunnable\n",
+    "\n",
+    "rag_app = RemoteRunnable(\"http://localhost:8001/multi-modal-rag\")\n",
+    "rag_app.invoke(\"How does the share of large-large AI results shift from 2012 to 2022?\")"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.9.16"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

templates/rag-multi-modal/rag_multi_modal/__init__.py

@@ -0,0 +1,3 @@
+from rag_multi_modal.chain import chain
+
+__all__ = ["chain"]

templates/rag-multi-modal/rag_multi_modal/chain.py

@@ -0,0 +1,94 @@
+# Load
+import os
+import uuid
+
+import chromadb
+import numpy as np
+from chromadb.utils.embedding_functions import OpenCLIPEmbeddingFunction
+from langchain.chat_models import ChatOpenAI
+from langchain.prompts import ChatPromptTemplate
+from langchain.schema.output_parser import StrOutputParser
+from langchain.schema.runnable import RunnablePassthrough
+from langchain.vectorstores import Chroma
+from PIL import Image as _PILImage
+from pydantic import BaseModel
+from unstructured.partition.pdf import partition_pdf
+
+# File
+path = "tests/ai_labs/"
+paper = "ai_labs.pdf"
+
+# Load and partition
+raw_pdf_elements = partition_pdf(
+    filename=path + paper,
+    extract_images_in_pdf=True,  # Extract images
+    infer_table_structure=True,  # Post processing to aggregate text into sections
+    chunking_strategy="by_title",
+    max_characters=4000,
+    new_after_n_chars=3800,
+    combine_text_under_n_chars=2000,
+    image_output_dir_path=path,
+)
+
+# Get texts and tables
+tables = []
+texts = []
+for element in raw_pdf_elements:
+    if "unstructured.documents.elements.Table" in str(type(element)):
+        tables.append(str(element))
+    elif "unstructured.documents.elements.CompositeElement" in str(type(element)):
+        texts.append(str(element))
+
+# Get images
+image_files = [f for f in os.listdir(path) if f.endswith(".jpg")]
+images = [np.array(_PILImage.open(path + f).convert("RGB")) for f in image_files]
+
+# Store in Chroma with multimodal embd
+## TO DO: Merge
+client = chromadb.Client()
+embedding_function = OpenCLIPEmbeddingFunction()
+collection = client.create_collection("mm_rag", embedding_function=embedding_function)
+
+image_ids = [str(uuid.uuid4()) for _ in images]
+collection.add(ids=image_ids, images=images)
+
+text_ids = [str(uuid.uuid4()) for _ in texts]
+collection.add(ids=text_ids, documents=texts)
+collection.get(include=["documents"])
+
+# Pass Chroma Client to LangChain
+vectorstore = Chroma(
+    client=client,
+    collection_name="mm_rag",
+    embedding_function=embedding_function,
+)
+
+retriever = vectorstore.as_retriever()
+
+# RAG
+# Prompt template
+template = """Answer the question based only on the following context, which can include text and tables:
+{context}
+Question: {question}
+"""  # noqa: E501
+prompt = ChatPromptTemplate.from_template(template)
+
+# LLM
+### placeholder ###
+model = ChatOpenAI(temperature=0, model="gpt-4v")
+
+# RAG pipeline
+chain = (
+    {"context": retriever, "question": RunnablePassthrough()}
+    | prompt
+    | model
+    | StrOutputParser()
+)
+
+
+# Add typing for input
+class Question(BaseModel):
+    __root__: str
+
+
+chain = chain.with_types(input_type=Question)

templates/rag-multi-modal/tests/Multi-Modal-Eval.csv

@@ -0,0 +1,10 @@
+Paper,Question,Answer
+wildfire.pdf,What is the difference in acres burned from Wildfires between 1993 and 2022?,"Acres burned from wildfires increases from ~2M in 1993 to ~8M in 2022, a ~4x increase."
+wildfire.pdf,What is the trend in the total number of wildfires between 1993 and 2022?,The total number of wildfires between 1993 and 2022 has oscillated up and down without a clear trend in either direction. The total number of fires in 1993 is about equal to the number in 2022.
+housing.pdf,How does housing compare to medical care in CPI weights?,Housing is 42% of CPI whereas medical care is 9%.
+housing.pdf,What’s the share of recreation in the CPI weights?,Recreation is 6% of CPI.
+ai_labs.pdf,How does the share of large-large AI results shift from 2012 to 2022?,The majority of results in 2012 came from academia whereas by 2022 research consortiums and academia have an equal share.
+ai_labs.pdf,When did research consortiums start to play a role in large scale AI results?,Research consortiums started to produce large scale AI results in 2021.
+nvda.pdf,What is the acceleration benefit from A100 on Physics HPC applications?,A100 achieves 1.9x and 2.1x acceleration on LAMMPS and Chroma for physics.
+bridewater.pdf,What is the trend in call center employment from 2000 to 2023?,"Call center employment rose significantly between around 2005 and around 2015. It has then falling since around 2015, most notably in the past 19 months."
+bridewater.pdf,What is the typical time lag between the first examples of a new technology and the resulting rise in productivity growth?,There is typically a multi-year lag between the fist demonstration of technology and their resulting impact on productivity growth: for electrification it was around 20 years and for PCs it was also around 20 years.