MultimodalQnA image query, pdf, and dynamic ports

comps/cores/proto/docarray.py

@@ -278,7 +278,7 @@ class GraphDoc(BaseDoc):
 
 
 class LVMDoc(BaseDoc):
-    image: str
+    image: Union[str, List[str]]
     prompt: str
     max_new_tokens: conint(ge=0, le=1024) = 512
     top_k: int = 10

comps/dataprep/multimodal/redis/langchain/README.md

@@ -5,6 +5,7 @@ This `dataprep` microservice accepts the following from the user and ingests the
 - Videos (mp4 files) and their transcripts (optional)
 - Images (gif, jpg, jpeg, and png files) and their captions (optional)
 - Audio (wav files)
+- PDFs (with text and images)
 
 ## 🚀1. Start Microservice with Python（Option 1）
 
@@ -111,18 +112,19 @@ docker container logs -f dataprep-multimodal-redis
 
 ## 🚀4. Consume Microservice
 
-Once this dataprep microservice is started, user can use the below commands to invoke the microservice to convert images and videos and their transcripts (optional) to embeddings and save to the Redis vector store.
+Once this dataprep microservice is started, user can use the below commands to invoke the microservice to convert images, videos, text, and PDF files to embeddings and save to the Redis vector store.
 
 This microservice provides 3 different ways for users to ingest files into Redis vector store corresponding to the 3 use cases.
 
 ### 4.1 Consume _ingest_with_text_ API
 
-**Use case:** This API is used when videos are accompanied by transcript files (`.vtt` format) or images are accompanied by text caption files (`.txt` format).
+**Use case:** This API is used for videos accompanied by transcript files (`.vtt` format), images accompanied by text caption files (`.txt` format), and PDF files containing a mix of text and images.
 
 **Important notes:**
 
 - Make sure the file paths after `files=@` are correct.
 - Every transcript or caption file's name must be identical to its corresponding video or image file's name (except their extension - .vtt goes with .mp4 and .txt goes with .jpg, .jpeg, .png, or .gif). For example, `video1.mp4` and `video1.vtt`. Otherwise, if `video1.vtt` is not included correctly in the API call, the microservice will return an error `No captions file video1.vtt found for video1.mp4`.
+- It is assumed that PDFs will contain at least one image. Each image in the file will be embedded along with the text that appears on the same page as the image.
 
 #### Single video-transcript pair upload
 
@@ -157,6 +159,7 @@ curl -X POST \
     -F "files=@./image1.txt" \
     -F "files=@./image2.jpg" \
     -F "files=@./image2.txt" \
+    -F "files=@./example.pdf" \
     http://localhost:6007/v1/ingest_with_text
 ```
 

comps/dataprep/multimodal/redis/langchain/prepare_videodoc_redis.py

@@ -1,13 +1,16 @@
 # Copyright (C) 2024 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
 
+import base64
+import json
 import os
 import shutil
 import time
 import uuid
 from pathlib import Path
 from typing import Any, Dict, Iterable, List, Optional, Type, Union
 
+import pymupdf
 from config import EMBED_MODEL, INDEX_NAME, INDEX_SCHEMA, LVM_ENDPOINT, REDIS_URL, WHISPER_MODEL
 from fastapi import File, HTTPException, UploadFile
 from langchain_community.utilities.redis import _array_to_buffer
@@ -301,7 +304,53 @@ def prepare_data_and_metadata_from_annotation(
     return text_list, image_list, metadatas
 
 
-def ingest_multimodal(videoname, data_folder, embeddings):
+def prepare_pdf_data_from_annotation(annotation, path_to_files, title):
+    """PDF data processing has some key differences from videos and images.
+
+    1. Neighboring transcripts are not currently considered relevant.
+       We are only taking the text located on the same page as the image.
+    2. The images within PDFs are indexed by page and image-within-page
+       indices, as opposed to a single frame index.
+    3. Instead of time of frame in ms, we return the PDF page index through
+       the pre-existing time_of_frame_ms metadata key to maintain compatibility.
+    """
+    text_list = []
+    image_list = []
+    metadatas = []
+    for item in annotation:
+        page_index = item["frame_no"]
+        image_index = item["sub_video_id"]
+        path_to_image = os.path.join(path_to_files, f"page{page_index}_image{image_index}.png")
+        caption_for_ingesting = item["caption"]
+        caption_for_inference = item["caption"]
+
+        pdf_id = item["video_id"]
+        b64_img_str = item["b64_img_str"]
+        embedding_type = "pair" if b64_img_str else "text"
+        source = item["video_name"]
+
+        text_list.append(caption_for_ingesting)
+
+        if b64_img_str:
+            image_list.append(path_to_image)
+
+        metadatas.append(
+            {
+                "content": caption_for_ingesting,
+                "b64_img_str": b64_img_str,
+                "video_id": pdf_id,
+                "source_video": source,
+                "time_of_frame_ms": page_index,  # For PDFs save the page number
+                "embedding_type": embedding_type,
+                "title": title,
+                "transcript_for_inference": caption_for_inference,
+            }
+        )
+
+    return text_list, image_list, metadatas
+
+
+def ingest_multimodal(filename, data_folder, embeddings, is_pdf=False):
     """Ingest text image pairs to Redis from the data/ directory that consists of frames and annotations."""
     data_folder = os.path.abspath(data_folder)
     annotation_file_path = os.path.join(data_folder, "annotations.json")
@@ -310,10 +359,15 @@ def ingest_multimodal(videoname, data_folder, embeddings):
     annotation = load_json_file(annotation_file_path)
 
     # prepare data to ingest
-    text_list, image_list, metadatas = prepare_data_and_metadata_from_annotation(annotation, path_to_frames, videoname)
+    if is_pdf:
+        text_list, image_list, metadatas = prepare_pdf_data_from_annotation(annotation, path_to_frames, filename)
+    else:
+        text_list, image_list, metadatas = prepare_data_and_metadata_from_annotation(
+            annotation, path_to_frames, filename
+        )
 
     MultimodalRedis.from_text_image_pairs_return_keys(
-        texts=[f"From {videoname}. " + text for text in text_list],
+        texts=[f"From {filename}. " + text for text in text_list],
         images=image_list,
         embedding=embeddings,
         metadatas=metadatas,
@@ -335,7 +389,10 @@ def drop_index(index_name, redis_url=REDIS_URL):
 
 
 @register_microservice(
-    name="opea_service@prepare_videodoc_redis", endpoint="/v1/generate_transcripts", host="0.0.0.0", port=6007
+    name="opea_service@prepare_videodoc_redis",
+    endpoint="/v1/generate_transcripts",
+    host="0.0.0.0",
+    port=int(os.getenv("DATAPREP_MMR_PORT", 6007)),
 )
 async def ingest_generate_transcripts(files: List[UploadFile] = File(None)):
     """Upload videos or audio files with speech, generate transcripts using whisper and ingest into redis."""
@@ -444,7 +501,10 @@ async def ingest_generate_transcripts(files: List[UploadFile] = File(None)):
 
 
 @register_microservice(
-    name="opea_service@prepare_videodoc_redis", endpoint="/v1/generate_captions", host="0.0.0.0", port=6007
+    name="opea_service@prepare_videodoc_redis",
+    endpoint="/v1/generate_captions",
+    host="0.0.0.0",
+    port=int(os.getenv("DATAPREP_MMR_PORT", 6007)),
 )
 async def ingest_generate_caption(files: List[UploadFile] = File(None)):
     """Upload images and videos without speech (only background music or no audio), generate captions using lvm microservice and ingest into redis."""
@@ -506,11 +566,11 @@ async def ingest_generate_caption(files: List[UploadFile] = File(None)):
     name="opea_service@prepare_videodoc_redis",
     endpoint="/v1/ingest_with_text",
     host="0.0.0.0",
-    port=6007,
+    port=int(os.getenv("DATAPREP_MMR_PORT", 6007)),
 )
 async def ingest_with_text(files: List[UploadFile] = File(None)):
     if files:
-        accepted_media_formats = [".mp4", ".png", ".jpg", ".jpeg", ".gif"]
+        accepted_media_formats = [".mp4", ".png", ".jpg", ".jpeg", ".gif", ".pdf"]
         # Create a lookup dictionary containing all media files
         matched_files = {f.filename: [f] for f in files if os.path.splitext(f.filename)[1] in accepted_media_formats}
         uploaded_files_map = {}
@@ -537,25 +597,25 @@ async def ingest_with_text(files: List[UploadFile] = File(None)):
             elif file_extension not in accepted_media_formats:
                 print(f"Skipping file {file.filename} because of unsupported format.")
 
-        # Check if every media file has a caption file
-        for media_file_name, file_pair in matched_files.items():
-            if len(file_pair) != 2:
+        # Check that every media file that is not a pdf has a caption file
+        for media_file_name, file_list in matched_files.items():
+            if len(file_list) != 2 and os.path.splitext(media_file_name)[1] != ".pdf":
                 raise HTTPException(status_code=400, detail=f"No caption file found for {media_file_name}")
 
         if len(matched_files.keys()) == 0:
             return HTTPException(
                 status_code=400,
-                detail="The uploaded files have unsupported formats. Please upload at least one video file (.mp4) with captions (.vtt) or one image (.png, .jpg, .jpeg, or .gif) with caption (.txt)",
+                detail="The uploaded files have unsupported formats. Please upload at least one video file (.mp4) with captions (.vtt) or one image (.png, .jpg, .jpeg, or .gif) with caption (.txt) or one .pdf file",
             )
 
         for media_file in matched_files:
             print(f"Processing file {media_file}")
+            file_name, file_extension = os.path.splitext(media_file)
 
             # Assign unique identifier to file
             file_id = generate_id()
 
             # Create file name by appending identifier
-            file_name, file_extension = os.path.splitext(media_file)
             media_file_name = f"{file_name}_{file_id}{file_extension}"
             media_dir_name = os.path.splitext(media_file_name)[0]
 
@@ -564,25 +624,72 @@ async def ingest_with_text(files: List[UploadFile] = File(None)):
                 shutil.copyfileobj(matched_files[media_file][0].file, f)
             uploaded_files_map[file_name] = media_file_name
 
-            # Save caption file in upload directory
-            caption_file_extension = os.path.splitext(matched_files[media_file][1].filename)[1]
-            caption_file = f"{media_dir_name}{caption_file_extension}"
-            with open(os.path.join(upload_folder, caption_file), "wb") as f:
-                shutil.copyfileobj(matched_files[media_file][1].file, f)
+            if file_extension == ".pdf":
+                # Set up location to store pdf images and text, reusing "frames" and "annotations" from video
+                output_dir = os.path.join(upload_folder, media_dir_name)
+                os.makedirs(output_dir, exist_ok=True)
+                os.makedirs(os.path.join(output_dir, "frames"), exist_ok=True)
+                doc = pymupdf.open(os.path.join(upload_folder, media_file_name))
+                annotations = []
+                for page_idx, page in enumerate(doc, start=1):
+                    text = page.get_text()
+                    images = page.get_images()
+                    for image_idx, image in enumerate(images, start=1):
+                        # Write image and caption file for each image found in pdf
+                        img_fname = f"page{page_idx}_image{image_idx}"
+                        img_fpath = os.path.join(output_dir, "frames", img_fname + ".png")
+                        pix = pymupdf.Pixmap(doc, image[0])  # create pixmap
+
+                        if pix.n - pix.alpha > 3:  # if CMYK, convert to RGB first
+                            pix = pymupdf.Pixmap(pymupdf.csRGB, pix)
+
+                        pix.save(img_fpath)  # pixmap to png
+                        pix = None
+
+                        # Convert image to base64 encoded string
+                        with open(img_fpath, "rb") as image2str:
+                            encoded_string = base64.b64encode(image2str.read())  # png to bytes
+
+                        decoded_string = encoded_string.decode()  # bytes to string
+
+                        # Create annotations file, reusing metadata keys from video
+                        annotations.append(
+                            {
+                                "video_id": file_id,
+                                "video_name": os.path.basename(os.path.join(upload_folder, media_file_name)),
+                                "b64_img_str": decoded_string,
+                                "caption": text,
+                                "time": 0.0,
+                                "frame_no": page_idx,
+                                "sub_video_id": image_idx,
+                            }
+                        )
+
+                with open(os.path.join(output_dir, "annotations.json"), "w") as f:
+                    json.dump(annotations, f)
+
+                # Ingest multimodal data into redis
+                ingest_multimodal(file_name, os.path.join(upload_folder, media_dir_name), embeddings, is_pdf=True)
+            else:
+                # Save caption file in upload directory
+                caption_file_extension = os.path.splitext(matched_files[media_file][1].filename)[1]
+                caption_file = f"{media_dir_name}{caption_file_extension}"
+                with open(os.path.join(upload_folder, caption_file), "wb") as f:
+                    shutil.copyfileobj(matched_files[media_file][1].file, f)
 
-            # Store frames and caption annotations in a new directory
-            extract_frames_and_annotations_from_transcripts(
-                file_id,
-                os.path.join(upload_folder, media_file_name),
-                os.path.join(upload_folder, caption_file),
-                os.path.join(upload_folder, media_dir_name),
-            )
+                # Store frames and caption annotations in a new directory
+                extract_frames_and_annotations_from_transcripts(
+                    file_id,
+                    os.path.join(upload_folder, media_file_name),
+                    os.path.join(upload_folder, caption_file),
+                    os.path.join(upload_folder, media_dir_name),
+                )
 
-            # Delete temporary caption file
-            os.remove(os.path.join(upload_folder, caption_file))
+                # Delete temporary caption file
+                os.remove(os.path.join(upload_folder, caption_file))
 
-            # Ingest multimodal data into redis
-            ingest_multimodal(file_name, os.path.join(upload_folder, media_dir_name), embeddings)
+                # Ingest multimodal data into redis
+                ingest_multimodal(file_name, os.path.join(upload_folder, media_dir_name), embeddings)
 
             # Delete temporary media directory containing frames and annotations
             shutil.rmtree(os.path.join(upload_folder, media_dir_name))
@@ -602,7 +709,10 @@ async def ingest_with_text(files: List[UploadFile] = File(None)):
 
 
 @register_microservice(
-    name="opea_service@prepare_videodoc_redis", endpoint="/v1/dataprep/get_files", host="0.0.0.0", port=6007
+    name="opea_service@prepare_videodoc_redis",
+    endpoint="/v1/dataprep/get_files",
+    host="0.0.0.0",
+    port=int(os.getenv("DATAPREP_MMR_PORT", 6007)),
 )
 async def rag_get_file_structure():
     """Returns list of names of uploaded videos saved on the server."""
@@ -616,7 +726,10 @@ async def rag_get_file_structure():
 
 
 @register_microservice(
-    name="opea_service@prepare_videodoc_redis", endpoint="/v1/dataprep/delete_files", host="0.0.0.0", port=6007
+    name="opea_service@prepare_videodoc_redis",
+    endpoint="/v1/dataprep/delete_files",
+    host="0.0.0.0",
+    port=int(os.getenv("DATAPREP_MMR_PORT", 6007)),
 )
 async def delete_files():
     """Delete all uploaded files along with redis index."""

comps/dataprep/multimodal/redis/langchain/requirements.txt

@@ -11,6 +11,7 @@ opentelemetry-sdk
 Pillow
 prometheus-fastapi-instrumentator
 pydantic
+pymupdf
 python-multipart
 redis
 shortuuid

comps/embeddings/src/integrations/multimodal_bridgetower.py

@@ -51,9 +51,13 @@ async def invoke(self, input: MultimodalDoc) -> EmbedMultimodalDoc:
             json["text"] = input.text
         elif isinstance(input, TextImageDoc):
             json["text"] = input.text.text
-            img_bytes = input.image.url.load_bytes()
-            base64_img = base64.b64encode(img_bytes).decode("utf-8")
-            json["img_b64_str"] = base64_img
+            if input.image.url:
+                img_bytes = input.image.url.load_bytes()
+                base64_img = base64.b64encode(img_bytes).decode("utf-8")
+            elif input.image.base64_image:
+                base64_img = input.image.base64_image
+            if base64_img:
+                json["img_b64_str"] = base64_img
         else:
             raise TypeError(
                 f"Unsupported input type: {type(input)}. "
@@ -71,6 +75,9 @@ async def invoke(self, input: MultimodalDoc) -> EmbedMultimodalDoc:
         elif isinstance(input, TextImageDoc):
             res = EmbedMultimodalDoc(text=input.text.text, url=input.image.url, embedding=embed_vector)
 
+            if base64_img:
+                res.base64_image = base64_img
+
         return res
 
     def check_health(self) -> bool:

comps/lvms/src/integrations/dependency/llava/README.md

@@ -1,6 +1,6 @@
 # LVM Microservice
 
-Visual Question and Answering is one of the multimodal tasks empowered by LVMs (Large Visual Models). This microservice supports visual Q&A by using LLaVA as the base large visual model. It accepts two inputs: a prompt and an image. It outputs the answer to the prompt about the image.
+Visual Question and Answering is one of the multimodal tasks empowered by LVMs (Large Visual Models). This microservice supports visual Q&A by using LLaVA as the base large visual model. It accepts two inputs: a prompt and images. It outputs the answer to the prompt about the images.
 
 ## 🚀1. Start Microservice with Python (Option 1)
 
@@ -73,7 +73,23 @@ docker run -p 8399:8399 --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_M
 
 #### 2.2.2 Test
 
+> Note: The `MAX_IMAGES` environment variable is used to specify the maximum number of images that will be sent from the LVM service to the LLaVA server.
+> If an image list longer than `MAX_IMAGES` is sent to the LVM server, a shortened image list will be sent to the LLaVA service. If the image list
+> needs to be shortened, the most recent images (the ones at the end of the list) are prioritized to send to the LLaVA service. Some LLaVA models have not
+> been trained with multiple images and may lead to inaccurate results. If `MAX_IMAGES` is not set, it will default to `1`.
+
 ```bash
+# Use curl/python
+
+# curl with an image and a prompt
+http_proxy="" curl http://localhost:9399/v1/lvm -XPOST -d '{"image": "iVBORw0KGgoAAAANSUhEUgAAAAoAAAAKCAYAAACNMs+9AAAAFUlEQVR42mP8/5+hnoEIwDiqkL4KAcT9GO0U4BxoAAAAAElFTkSuQmCC", "prompt":"What is this?"}' -H 'Content-Type: application/json'
+
+# curl with multiple images and a prompt (Note that depending on your MAX_IMAGES value, both images may not be sent to the LLaVA model)
+http_proxy="" curl http://localhost:9399/v1/lvm -XPOST -d '{"image": ["iVBORw0KGgoAAAANSUhEUgAAAAoAAAAKCAYAAACNMs+9AAAAFUlEQVR42mNkYPhfz0AEYBxVSF+FAP5FDvcfRYWgAAAAAElFTkSuQmCC", "iVBORw0KGgoAAAANSUhEUgAAAAoAAAAKCAYAAACNMs+9AAAAFUlEQVR42mNk+M9Qz0AEYBxVSF+FAAhKDveksOjmAAAAAElFTkSuQmCC"], "prompt":"What is in these images?"}' -H 'Content-Type: application/json'
+
+# curl with a prompt only (no image)
+http_proxy="" curl http://localhost:9399/v1/lvm -XPOST -d '{"image": "", "prompt":"What is deep learning?"}' -H 'Content-Type: application/json'
+
 # Test
 python check_llava_server.py
 ```