Retriever and lvm update for multimodal rag on videos (#606)

* updates

Signed-off-by: Tiep Le <tiep.le@intel.com>

* cosmetic

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* update redis schema

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* update multimodal config and docker compose retriever

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* update requirements

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* update retriever redis

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* multimodal retriever implementation

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* test for multimodal retriever

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* include prompt preparation for multimodal rag on videos application

Signed-off-by: sjagtap1803 <siddhant.jagtap@intel.com>

* fix template

Signed-off-by: sjagtap1803 <siddhant.jagtap@intel.com>

* add test for llava for mm_rag_on_videos

Signed-off-by: sjagtap1803 <siddhant.jagtap@intel.com>

* update test

Signed-off-by: sjagtap1803 <siddhant.jagtap@intel.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix index not found

Signed-off-by: sjagtap1803 <siddhant.jagtap@intel.com>

* add LVMSearchedMultimodalDoc

Signed-off-by: sjagtap1803 <siddhant.jagtap@intel.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* remove INDEX_SCHEMA

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* revise folder structure to comps/retrievers/langchain/redis_multimodal

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* update test

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* change port of redis to resolve CI test

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* update test

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

* update lvms test

Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>

---------

Signed-off-by: Tiep Le <tiep.le@intel.com>
Signed-off-by: siddhivelankar23 <siddhi.velankar@intel.com>
Signed-off-by: sjagtap1803 <siddhant.jagtap@intel.com>
Co-authored-by: siddhivelankar23 <siddhi.velankar@intel.com>
Co-authored-by: sjagtap1803 <siddhant.jagtap@intel.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

comps/__init__.py

@@ -13,6 +13,7 @@
     LLMParamsDoc,
     SearchedDoc,
     SearchedMultimodalDoc,
+    LVMSearchedMultimodalDoc,
     RerankedDoc,
     TextDoc,
     RAGASParams,

comps/cores/proto/docarray.py

@@ -107,6 +107,24 @@ class SearchedMultimodalDoc(SearchedDoc):
     metadata: List[Dict[str, Any]]
 
 
+class LVMSearchedMultimodalDoc(SearchedMultimodalDoc):
+    max_new_tokens: conint(ge=0, le=1024) = 512
+    top_k: int = 10
+    top_p: float = 0.95
+    typical_p: float = 0.95
+    temperature: float = 0.01
+    streaming: bool = False
+    repetition_penalty: float = 1.03
+    chat_template: Optional[str] = Field(
+        default=None,
+        description=(
+            "A template to use for this conversion. "
+            "If this is not passed, the model's default chat template will be "
+            "used instead. We recommend that the template contains {context} and {question} for multimodal-rag on videos."
+        ),
+    )
+
+
 class GeneratedDoc(BaseDoc):
     text: str
     prompt: str

comps/lvms/lvm.py

@@ -5,12 +5,16 @@
 import json
 import os
 import time
+from typing import Union
 
 import requests
+from langchain_core.prompts import PromptTemplate
+from template import ChatTemplate
 
 from comps import (
     CustomLogger,
     LVMDoc,
+    LVMSearchedMultimodalDoc,
     ServiceType,
     TextDoc,
     opea_microservices,
@@ -29,20 +33,41 @@
     endpoint="/v1/lvm",
     host="0.0.0.0",
     port=9399,
-    input_datatype=LVMDoc,
-    output_datatype=TextDoc,
 )
 @register_statistics(names=["opea_service@lvm"])
-async def lvm(request: LVMDoc):
+async def lvm(request: Union[LVMDoc, LVMSearchedMultimodalDoc]) -> TextDoc:
     if logflag:
         logger.info(request)
     start = time.time()
-    img_b64_str = request.image
-    prompt = request.prompt
-    max_new_tokens = request.max_new_tokens
+    if isinstance(request, LVMSearchedMultimodalDoc):
+        if logflag:
+            logger.info("[LVMSearchedMultimodalDoc ] input from retriever microservice")
+        retrieved_metadatas = request.metadata
+        img_b64_str = retrieved_metadatas[0]["b64_img_str"]
+        initial_query = request.initial_query
+        context = retrieved_metadatas[0]["transcript_for_inference"]
+        prompt = initial_query
+        if request.chat_template is None:
+            prompt = ChatTemplate.generate_multimodal_rag_on_videos_prompt(initial_query, context)
+        else:
+            prompt_template = PromptTemplate.from_template(request.chat_template)
+            input_variables = prompt_template.input_variables
+            if sorted(input_variables) == ["context", "question"]:
+                prompt = prompt_template.format(question=initial_query, context=context)
+            else:
+                logger.info(
+                    f"[ LVMSearchedMultimodalDoc ] {prompt_template} not used, we only support 2 input variables ['question', 'context']"
+                )
+        max_new_tokens = request.max_new_tokens
+        if logflag:
+            logger.info(f"prompt generated for [LVMSearchedMultimodalDoc ] input from retriever microservice: {prompt}")
 
-    inputs = {"img_b64_str": img_b64_str, "prompt": prompt, "max_new_tokens": max_new_tokens}
+    else:
+        img_b64_str = request.image
+        prompt = request.prompt
+        max_new_tokens = request.max_new_tokens
 
+    inputs = {"img_b64_str": img_b64_str, "prompt": prompt, "max_new_tokens": max_new_tokens}
     # forward to the LLaVA server
     response = requests.post(url=f"{lvm_endpoint}/generate", data=json.dumps(inputs), proxies={"http": None})
 

comps/lvms/requirements.txt

@@ -2,6 +2,7 @@ datasets
 docarray[full]
 fastapi
 huggingface_hub
+langchain-core
 opentelemetry-api
 opentelemetry-exporter-otlp
 opentelemetry-sdk

comps/lvms/template.py

@@ -0,0 +1,10 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+
+class ChatTemplate:
+
+    @staticmethod
+    def generate_multimodal_rag_on_videos_prompt(question: str, context: str):
+        template = """The transcript associated with the image is '{context}'. {question}"""
+        return template.format(context=context, question=question)

comps/retrievers/langchain/redis_multimodal/README.md

@@ -0,0 +1,123 @@
+# Retriever Microservice
+
+This retriever microservice is a highly efficient search service designed for handling and retrieving embedding vectors from multimodal data. It operates by receiving an embedding vector as input and conducting a similarity search against vectors stored in a VectorDB database. Users must specify the VectorDB's URL and the index name, and the service searches within that index to find documents with the highest similarity to the input vector.
+
+The service primarily utilizes similarity measures in vector space to rapidly retrieve contentually similar documents. The vector-based retrieval approach is particularly suited for handling large datasets, offering fast and accurate search results that significantly enhance the efficiency and quality of information retrieval.
+
+Overall, this microservice provides robust backend support for applications requiring efficient similarity searches, playing a vital role in scenarios such as recommendation systems, information retrieval, or any other context where precise measurement of document similarity is crucial.
+
+## 🚀1. Start Microservice with Python (Option 1)
+
+To start the retriever microservice, you must first install the required python packages.
+
+### 1.1 Install Requirements
+
+```bash
+pip install -r requirements.txt
+```
+
+### 1.2 Setup VectorDB Service
+
+You need to setup your own VectorDB service (Redis in this example), and ingest your knowledge documents into the vector database.
+
+As for Redis, you could start a docker container using the following commands.
+Remember to ingest data into it manually.
+
+```bash
+docker run -d --name="redis-vector-db" -p 6379:6379 -p 8001:8001 redis/redis-stack:7.2.0-v9
+```
+
+### 1.3 Ingest images or video
+
+Upload a video or images using the dataprep microservice, instructions can be found [here](https://github.com/opea-project/GenAIComps/tree/main/comps/dataprep/redis/multimodal_langchain/README.md).
+
+### 1.4 Start Retriever Service
+
+```bash
+python retriever_redis.py
+```
+
+## 🚀2. Start Microservice with Docker (Option 2)
+
+### 2.1 Setup Environment Variables
+
+```bash
+export your_ip=$(hostname -I | awk '{print $1}')
+export REDIS_URL="redis://${your_ip}:6379"
+export INDEX_NAME=${your_index_name}
+```
+
+### 2.2 Build Docker Image
+
+```bash
+cd ../../../../
+docker build -t opea/multimodal-retriever-redis:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/retrievers/langchain/redis_multimodal/docker/Dockerfile .
+```
+
+To start a docker container, you have two options:
+
+- A. Run Docker with CLI
+- B. Run Docker with Docker Compose
+
+You can choose one as needed.
+
+### 2.3 Run Docker with CLI (Option A)
+
+```bash
+docker run -d --name="multimodal-retriever-redis-server" -p 7000:7000 --ipc=host -e http_proxy=$http_proxy -e https_proxy=$https_proxy -e REDIS_URL=$REDIS_URL -e INDEX_NAME=$INDEX_NAME opea/multimodal-retriever-redis:latest
+```
+
+### 2.4 Run Docker with Docker Compose (Option B)
+
+```bash
+cd docker
+docker compose -f docker_compose_retriever.yaml up -d
+```
+
+## 🚀3. Consume Retriever Service
+
+### 3.1 Consume Embedding Service
+
+To consume the Retriever Microservice, you can generate a mock embedding vector of length 512 with Python.
+
+```bash
+your_embedding=$(python -c "import random; embedding = [random.uniform(-1, 1) for _ in range(512)]; print(embedding)")
+curl http://${your_ip}:7000/v1/multimodal_retrieval \
+  -X POST \
+  -d "{\"text\":\"What is the revenue of Nike in 2023?\",\"embedding\":${your_embedding}}" \
+  -H 'Content-Type: application/json'
+```
+
+You can set the parameters for the retriever.
+
+```bash
+your_embedding=$(python -c "import random; embedding = [random.uniform(-1, 1) for _ in range(512)]; print(embedding)")
+curl http://localhost:7000/v1/multimodal_retrieval \
+  -X POST \
+  -d "{\"text\":\"What is the revenue of Nike in 2023?\",\"embedding\":${your_embedding},\"search_type\":\"similarity\", \"k\":4}" \
+  -H 'Content-Type: application/json'
+```
+
+```bash
+your_embedding=$(python -c "import random; embedding = [random.uniform(-1, 1) for _ in range(512)]; print(embedding)")
+curl http://localhost:7000/v1/multimodal_retrieval \
+  -X POST \
+  -d "{\"text\":\"What is the revenue of Nike in 2023?\",\"embedding\":${your_embedding},\"search_type\":\"similarity_distance_threshold\", \"k\":4, \"distance_threshold\":1.0}" \
+  -H 'Content-Type: application/json'
+```
+
+```bash
+your_embedding=$(python -c "import random; embedding = [random.uniform(-1, 1) for _ in range(512)]; print(embedding)")
+curl http://localhost:7000/v1/multimodal_retrieval \
+  -X POST \
+  -d "{\"text\":\"What is the revenue of Nike in 2023?\",\"embedding\":${your_embedding},\"search_type\":\"similarity_score_threshold\", \"k\":4, \"score_threshold\":0.2}" \
+  -H 'Content-Type: application/json'
+```
+
+```bash
+your_embedding=$(python -c "import random; embedding = [random.uniform(-1, 1) for _ in range(512)]; print(embedding)")
+curl http://localhost:7000/v1/multimodal_retrieval \
+  -X POST \
+  -d "{\"text\":\"What is the revenue of Nike in 2023?\",\"embedding\":${your_embedding},\"search_type\":\"mmr\", \"k\":4, \"fetch_k\":20, \"lambda_mult\":0.5}" \
+  -H 'Content-Type: application/json'
+```

comps/retrievers/langchain/redis_multimodal/__init__.py

@@ -0,0 +1,2 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0

comps/retrievers/langchain/redis_multimodal/docker/Dockerfile

@@ -0,0 +1,29 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+FROM langchain/langchain:latest
+
+ARG ARCH="cpu"
+
+RUN apt-get update -y && apt-get install -y --no-install-recommends --fix-missing \
+    libgl1-mesa-glx \
+    libjemalloc-dev \
+    vim
+
+RUN useradd -m -s /bin/bash user && \
+    mkdir -p /home/user && \
+    chown -R user /home/user/
+
+COPY comps /home/user/comps
+
+USER user
+
+RUN pip install --no-cache-dir --upgrade pip && \
+    if [ ${ARCH} = "cpu" ]; then pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu; fi && \
+    pip install --no-cache-dir -r /home/user/comps/retrievers/langchain/redis_multimodal/requirements.txt
+
+ENV PYTHONPATH=$PYTHONPATH:/home/user
+
+WORKDIR /home/user/comps/retrievers/langchain/redis_multimodal
+
+ENTRYPOINT ["python", "retriever_redis.py"]

comps/retrievers/langchain/redis_multimodal/docker/docker_compose_retriever.yaml

@@ -0,0 +1,23 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+version: "1.0"
+
+services:
+  retriever:
+    image: opea/multimodal-retriever-redis:latest
+    container_name: multimodal-retriever-redis-server
+    ports:
+      - "7000:7000"
+    ipc: host
+    environment:
+      no_proxy: ${no_proxy}
+      http_proxy: ${http_proxy}
+      https_proxy: ${https_proxy}
+      REDIS_URL: ${REDIS_URL}
+      INDEX_NAME: ${INDEX_NAME}
+    restart: unless-stopped
+
+networks:
+  default:
+    driver: bridge

comps/retrievers/langchain/redis_multimodal/multimodal_config.py

@@ -0,0 +1,77 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import os
+
+current_file_path = os.path.abspath(__file__)
+parent_dir = os.path.dirname(current_file_path)
+
+
+def get_boolean_env_var(var_name, default_value=False):
+    """Retrieve the boolean value of an environment variable.
+
+    Args:
+    var_name (str): The name of the environment variable to retrieve.
+    default_value (bool): The default value to return if the variable
+    is not found.
+    Returns:
+    bool: The value of the environment variable, interpreted as a boolean.
+    """
+    true_values = {"true", "1", "t", "y", "yes"}
+    false_values = {"false", "0", "f", "n", "no"}
+
+    # Retrieve the environment variable's value
+    value = os.getenv(var_name, "").lower()
+
+    # Decide the boolean value based on the content of the string
+    if value in true_values:
+        return True
+    elif value in false_values:
+        return False
+    else:
+        return default_value
+
+
+# Check for openai API key
+# if "OPENAI_API_KEY" not in os.environ:
+#    raise Exception("Must provide an OPENAI_API_KEY as an env var.")
+
+
+# Whether or not to enable langchain debugging
+DEBUG = get_boolean_env_var("DEBUG", False)
+# Set DEBUG env var to "true" if you wish to enable LC debugging module
+if DEBUG:
+    import langchain
+
+    langchain.debug = True
+
+
+# Embedding model
+EMBED_MODEL = os.getenv("EMBED_MODEL", "BridgeTower/bridgetower-large-itm-mlm-itc")
+
+# Redis Connection Information
+REDIS_HOST = os.getenv("REDIS_HOST", "localhost")
+REDIS_PORT = int(os.getenv("REDIS_PORT", 6379))
+
+
+def format_redis_conn_from_env():
+    redis_url = os.getenv("REDIS_URL", None)
+    if redis_url:
+        return redis_url
+    else:
+        using_ssl = get_boolean_env_var("REDIS_SSL", False)
+        start = "rediss://" if using_ssl else "redis://"
+
+        # if using RBAC
+        password = os.getenv("REDIS_PASSWORD", None)
+        username = os.getenv("REDIS_USERNAME", "default")
+        if password is not None:
+            start += f"{username}:{password}@"
+
+        return start + f"{REDIS_HOST}:{REDIS_PORT}"
+
+
+REDIS_URL = format_redis_conn_from_env()
+
+# Vector Index Configuration
+INDEX_NAME = os.getenv("INDEX_NAME", "test-index")

comps/retrievers/langchain/redis_multimodal/requirements.txt

@@ -0,0 +1,11 @@
+docarray[full]
+fastapi
+langchain_community
+opentelemetry-api
+opentelemetry-exporter-otlp
+opentelemetry-sdk
+prometheus-fastapi-instrumentator
+redis
+shortuuid
+transformers
+uvicorn