Gemini multi-modal RAG template (#14678)

![Screenshot 2023-12-13 at 12 53 39
PM](https://github.com/langchain-ai/langchain/assets/122662504/a6bc3b0b-f177-4367-b9c8-b8862c847026)

templates/rag-gemini-multi-modal/.gitignore

@@ -0,0 +1 @@
+docs/img_*.jpg

templates/rag-gemini-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-gemini-multi-modal/README.md

@@ -0,0 +1,106 @@
+
+# rag-gemini-multi-modal
+
+Presentations (slide decks, etc) contain visual content that challenges conventional RAG.
+
+Multi-modal LLMs unlock new ways to build apps over visual content like presentations.
+ 
+This template performs multi-modal RAG using Chroma with multi-modal OpenCLIP embeddings and [Google Gemini](https://deepmind.google/technologies/gemini/#introduction).
+
+## Input
+
+Supply a slide deck as pdf in the `/docs` directory. 
+
+Create your vectorstore with: 
+
+```
+poetry install
+python ingest.py
+```
+
+## Embeddings
+
+This template will use [OpenCLIP](https://github.com/mlfoundations/open_clip) multi-modal embeddings.
+
+You can select different options (see results [here](https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_results.csv)).
+
+The first time you run the app, it will automatically download the multimodal embedding model.
+
+By default, LangChain will use an embedding model with reasonably strong performance, `ViT-H-14`.
+
+You can choose alternative `OpenCLIPEmbeddings` models in `ingest.py`:
+```
+vectorstore_mmembd = Chroma(
+    collection_name="multi-modal-rag",
+    persist_directory=str(re_vectorstore_path),
+    embedding_function=OpenCLIPEmbeddings(
+        model_name="ViT-H-14", checkpoint="laion2b_s32b_b79k"
+    ),
+)
+```
+
+## LLM
+
+The app will retrieve images using multi-modal embeddings, and pass them to Google Gemini.
+
+## Environment Setup
+
+Set the `GOOGLE_API_KEY` environment variable to access Gemini.
+
+## 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 rag-gemini-multi-modal
+```
+
+If you want to add this to an existing project, you can just run:
+
+```shell
+langchain app add rag-gemini-multi-modal
+```
+
+And add the following code to your `server.py` file:
+```python
+from rag_gemini_multi_modal import chain as rag_gemini_multi_modal_chain
+
+add_routes(app, rag_gemini_multi_modal_chain, path="/rag-gemini-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-gemini-multi-modal/playground](http://127.0.0.1:8000/rag-gemini-multi-modal/playground)  
+
+We can access the template from code with:
+
+```python
+from langserve.client import RemoteRunnable
+
+runnable = RemoteRunnable("http://localhost:8000/rag-gemini-multi-modal")
+```

templates/rag-gemini-multi-modal/ingest.py

@@ -0,0 +1,58 @@
+import os
+from pathlib import Path
+
+import pypdfium2 as pdfium
+from langchain.vectorstores import Chroma
+from langchain_experimental.open_clip import OpenCLIPEmbeddings
+
+
+def get_images_from_pdf(pdf_path, img_dump_path):
+    """
+    Extract images from each page of a PDF document and save as JPEG files.
+
+    :param pdf_path: A string representing the path to the PDF file.
+    :param img_dump_path: A string representing the path to dummp images.
+    """
+    pdf = pdfium.PdfDocument(pdf_path)
+    n_pages = len(pdf)
+    for page_number in range(n_pages):
+        page = pdf.get_page(page_number)
+        bitmap = page.render(scale=1, rotation=0, crop=(0, 0, 0, 0))
+        pil_image = bitmap.to_pil()
+        pil_image.save(f"{img_dump_path}/img_{page_number + 1}.jpg", format="JPEG")
+
+
+# Load PDF
+doc_path = Path(__file__).parent / "docs/DDOG_Q3_earnings_deck.pdf"
+img_dump_path = Path(__file__).parent / "docs/"
+rel_doc_path = doc_path.relative_to(Path.cwd())
+rel_img_dump_path = img_dump_path.relative_to(Path.cwd())
+print("pdf index")
+pil_images = get_images_from_pdf(rel_doc_path, rel_img_dump_path)
+print("done")
+vectorstore = Path(__file__).parent / "chroma_db_multi_modal"
+re_vectorstore_path = vectorstore.relative_to(Path.cwd())
+
+# Load embedding function
+print("Loading embedding function")
+embedding = OpenCLIPEmbeddings(model_name="ViT-H-14", checkpoint="laion2b_s32b_b79k")
+
+# Create chroma
+vectorstore_mmembd = Chroma(
+    collection_name="multi-modal-rag",
+    persist_directory=str(Path(__file__).parent / "chroma_db_multi_modal"),
+    embedding_function=embedding,
+)
+
+# Get image URIs
+image_uris = sorted(
+    [
+        os.path.join(rel_img_dump_path, image_name)
+        for image_name in os.listdir(rel_img_dump_path)
+        if image_name.endswith(".jpg")
+    ]
+)
+
+# Add images
+print("Embedding images")
+vectorstore_mmembd.add_images(uris=image_uris)

templates/rag-gemini-multi-modal/pyproject.toml

@@ -0,0 +1,39 @@
+[tool.poetry]
+name = "rag-gemini-multi-modal"
+version = "0.1.0"
+description = "Multi-modal RAG using Gemini and OpenCLIP embeddings"
+authors = [
+    "Lance Martin <lance@langchain.dev>",
+]
+readme = "README.md"
+
+[tool.poetry.dependencies]
+python = ">=3.9,<4.0"
+langchain = ">=0.0.350"
+openai = "<2"
+tiktoken = ">=0.5.1"
+chromadb = ">=0.4.14"
+open-clip-torch = ">=2.23.0"
+torch = ">=2.1.0"
+pypdfium2 = ">=4.20.0"
+langchain-experimental = "^0.0.43"
+langchain-google-genai = ">=0.0.1"
+
+[tool.poetry.group.dev.dependencies]
+langchain-cli = ">=0.0.15"
+
+[tool.langserve]
+export_module = "rag_gemini_multi_modal"
+export_attr = "chain"
+
+[tool.templates-hub]
+use-case = "rag"
+author = "LangChain"
+integrations = ["OpenAI", "Chroma"]
+tags = ["vectordbs"]
+
+[build-system]
+requires = [
+    "poetry-core",
+]
+build-backend = "poetry.core.masonry.api"

templates/rag-gemini-multi-modal/rag_gemini_multi_modal.ipynb

@@ -0,0 +1,52 @@
+{
+ "cells": [
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "id": "681a5d1e",
+   "metadata": {},
+   "source": [
+    "## Run Template\n",
+    "\n",
+    "In `server.py`, set -\n",
+    "```\n",
+    "add_routes(app, chain_rag_conv, path=\"/rag-gemini-multi-modal\")\n",
+    "```"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d774be2a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langserve.client import RemoteRunnable\n",
+    "\n",
+    "rag_app = RemoteRunnable(\"http://localhost:8001/rag-gemini-multi-modal\")\n",
+    "rag_app.invoke(\"What is the projected TAM for observability expected for each year through 2026?\")"
+   ]
+  }
+ ],
+ "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-gemini-multi-modal/rag_gemini_multi_modal/__init__.py

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

templates/rag-gemini-multi-modal/rag_gemini_multi_modal/chain.py

@@ -0,0 +1,122 @@
+import base64
+import io
+from pathlib import Path
+
+from langchain.vectorstores import Chroma
+from langchain_core.documents import Document
+from langchain_core.messages import HumanMessage
+from langchain_core.output_parsers import StrOutputParser
+from langchain_core.pydantic_v1 import BaseModel
+from langchain_core.runnables import RunnableLambda, RunnablePassthrough
+from langchain_experimental.open_clip import OpenCLIPEmbeddings
+from langchain_google_genai import ChatGoogleGenerativeAI
+from PIL import Image
+
+
+def resize_base64_image(base64_string, size=(128, 128)):
+    """
+    Resize an image encoded as a Base64 string.
+
+    :param base64_string: A Base64 encoded string of the image to be resized.
+    :param size: A tuple representing the new size (width, height) for the image.
+    :return: A Base64 encoded string of the resized image.
+    """
+    img_data = base64.b64decode(base64_string)
+    img = Image.open(io.BytesIO(img_data))
+    resized_img = img.resize(size, Image.LANCZOS)
+    buffered = io.BytesIO()
+    resized_img.save(buffered, format=img.format)
+    return base64.b64encode(buffered.getvalue()).decode("utf-8")
+
+
+def get_resized_images(docs):
+    """
+    Resize images from base64-encoded strings.
+
+    :param docs: A list of base64-encoded image to be resized.
+    :return: Dict containing a list of resized base64-encoded strings.
+    """
+    b64_images = []
+    for doc in docs:
+        if isinstance(doc, Document):
+            doc = doc.page_content
+        resized_image = resize_base64_image(doc, size=(1280, 720))
+        b64_images.append(resized_image)
+    return {"images": b64_images}
+
+
+def img_prompt_func(data_dict, num_images=2):
+    """
+    Gemini prompt for image analysis.
+
+    :param data_dict: A dict with images and a user-provided question.
+    :param num_images: Number of images to include in the prompt.
+    :return: A list containing message objects for each image and the text prompt.
+    """
+    messages = []
+    if data_dict["context"]["images"]:
+        for image in data_dict["context"]["images"][:num_images]:
+            image_message = {
+                "type": "image_url",
+                "image_url": {"url": f"data:image/jpeg;base64,{image}"},
+            }
+            messages.append(image_message)
+    text_message = {
+        "type": "text",
+        "text": (
+            "You are an analyst tasked with answering questions about visual content.\n"
+            "You will be give a set of image(s) from a slide deck / presentation.\n"
+            "Use this information to answer the user question. \n"
+            f"User-provided question: {data_dict['question']}\n\n"
+        ),
+    }
+    messages.append(text_message)
+    return [HumanMessage(content=messages)]
+
+
+def multi_modal_rag_chain(retriever):
+    """
+    Multi-modal RAG chain,
+
+    :param retriever: A function that retrieves the necessary context for the model.
+    :return: A chain of functions representing the multi-modal RAG process.
+    """
+    # Initialize the multi-modal Large Language Model with specific parameters
+    model = ChatGoogleGenerativeAI(model="gemini-pro-vision")
+
+    # Define the RAG pipeline
+    chain = (
+        {
+            "context": retriever | RunnableLambda(get_resized_images),
+            "question": RunnablePassthrough(),
+        }
+        | RunnableLambda(img_prompt_func)
+        | model
+        | StrOutputParser()
+    )
+
+    return chain
+
+
+# Load chroma
+vectorstore_mmembd = Chroma(
+    collection_name="multi-modal-rag",
+    persist_directory=str(Path(__file__).parent.parent / "chroma_db_multi_modal"),
+    embedding_function=OpenCLIPEmbeddings(
+        model_name="ViT-H-14", checkpoint="laion2b_s32b_b79k"
+    ),
+)
+
+# Make retriever
+retriever_mmembd = vectorstore_mmembd.as_retriever()
+
+# Create RAG chain
+chain = multi_modal_rag_chain(retriever_mmembd)
+
+
+# Add typing for input
+class Question(BaseModel):
+    __root__: str
+
+
+chain = chain.with_types(input_type=Question)