book-mentat

Considering how to analyse book collections, Large Language Model style

Motivation

• Get all the good spaceship names
• Finding old stories you have forgotten
• References found
• State of the art models for a current mineral exploration problem

Problem

Library

• Consider an extensive ebook library that is decades old

Formats

- text
- html
- rtf
- prc
- mobipocket
- epub
- pdf
- cbz / cbr

Origin

- Not all digital native
    - Failure OCR - what is open source state of the art
    - If on windows - robocopy multithreaded is useful robocopy "C:\Users\bananasplits\OneDrive\Calibre Books" "D:\books\Calibre Books" /MIR /MT:16

Type

• Fiction
• Non-fiction
• Games
• Academic
  • Papers etc
• Code
  • Many languages

Modality

• Do you want covers?
• Diagrams from non-fiction
• Pictures [or comic strips] from games
• Actual comics
  • 2000 AD, Image, Humble Bundles have CBR/CBZ possibilities

Current State

• Calibre https://calibre-ebook.com/

Prior Art

epub

• https://huggingface.co/learn/cookbook/en/rag_llamaindex_librarian
• https://github.com/huggingface/cookbook/blob/main/notebooks/en/rag_llamaindex_librarian.ipynb
• https://www.bitsgalore.org/2023/03/09/extracting-text-from-epub-files-in-python
  • https://github.com/RichardScottOZ/textExtractDemo

comics

• https://linnk.ai/insight/comics-analysis/multimodal-transformer-for-comics-text-cloze-enhancing-narrative-understanding-in-comics-analysis-f7rPlZEc/
  • paper https://arxiv.org/pdf/2403.03719

Approach

• Clearly will need some sort of Retrieval Augmented Generation

Data

Calibre Libraries

• Currently a folder with metadata, db and a folder per 'author' - which is not important, but is the structure

Technology

Storage

• Is in memory ok if run on a decent server
• A books library example

.azw: 12
.azw3: 234
.azw4: 2
.db: 2
.doc: 1
.docx: 1
.epub: 1567
.htmlz: 6
.jpg: 9829
.json: 1
.lit: 2
.lrf: 2
.mobi: 10054
.opf: 10891
.original_epub: 20
.pdb: 9
.pdf: 102
.prc: 315
.rar: 18
.rtf: 9
.txt: 28
.zip: 17

Calibre

• use this for conversion
  • ubuntu has an apt package

ollama

• install https://ollama.com/download/linux
  • curl -fsSL https://ollama.com/install.sh | sh
• models: https://ollama.com/library
• operation ollama/ollama#707

>>> Creating ollama user...
>>> Adding ollama user to render group...
>>> Adding ollama user to video group...
>>> Adding current user to ollama group...
>>> Creating ollama systemd service...
>>> Enabling and starting ollama service...
Created symlink /etc/systemd/system/default.target.wants/ollama.service → /etc/systemd/system/ollama.service.
>>> The Ollama API is now available at 127.0.0.1:11434.
>>> Install complete. Run "ollama" from the command line.
WARNING: No NVIDIA/AMD GPU detected. Ollama will run in CPU-only mode.

Llamaindex

Simple Directory Reader

• https://docs.llamaindex.ai/en/v0.9.48/examples/data_connectors/simple_directory_reader.html
• Readers https://github.com/run-llama/llama_index/tree/main/llama-index-integrations/readers/llama-index-readers-file
• Vector Stores https://docs.llamaindex.ai/en/stable/module_guides/storing/vector_stores/

Reranker

• https://docs.llamaindex.ai/en/stable/examples/node_postprocessor/LLMReranker-Gatsby/

Chroma

• https://docs.trychroma.com/
  • is multimodal
• installation https://docs.trychroma.com/getting-started
• vector store installation pip install llama-index-vector-stores-chroma
• embedding speed https://www.youtube.com/watch?v=7FvdwwvqrD4&ab_channel=JohnnyCode
• examples -
  • https://thenewstack.io/exploring-chroma-the-open-source-vector-database-for-llms/

Tokenisers

• ?

Models

• Text

• Multimodal - Llava?

• Model Storage ollama/ollama#1737

• https://ollama.com/library

• ollama pull llama2

• ollama run llama3

• etc

Ollama

Errors

bind address already in use

• ollama/ollama#707
• Hey all, not seeing ollama in the output of lsof could be a permissions issue. When you install ollama on linux via the install script it creates a service user for the background process. You may need to stop the process via systemctl in that case.

Here is some troubleshooting steps that will hopefully help:

Stop the background service:

sudo systemctl stop ollama

ollama serve  #to restart

Run lsof as sudo to rule out permissions issues:

sudo lsof -i :11434

Embeddings

• multi GPU approach
• https://github.com/UKPLab/sentence-transformers/blob/master/examples/applications/computing-embeddings/computing_embeddings_multi_gpu.py
• Chroma will do multimodal

Tips

• https://thenewstack.io/exploring-chroma-the-open-source-vector-database-for-llms/
• sqlite3 linux package can be used to check pragma basics etc.

Layout

• discussion https://www.reddit.com/r/LocalLLaMA/comments/1brk3qo/pdf_to_json_for_rag_through_multimodal_models/
• Llava?
• layoutlm https://huggingface.co/docs/transformers/en/model_doc/layoutlm
• grobid https://grobid.readthedocs.io/en/latest/Introduction/
• LlamaParse https://medium.com/@salujav4/parsing-pdfs-text-image-and-tables-for-rag-based-applications-using-llamaparse-llamaindex-0f4c5ed50fb7
  • https://colab.research.google.com/drive/1aUPywCH92XLNpdjkmXz3ff8H-QnT2JHZ?usp=sharing
• Unstructured https://github.com/Unstructured-IO

Tables

• https://github.com/parsee-ai/parsee-pdf-reader
• TableTransformer

PDF

• PyMuPDF
• PDFMiner.six
• Camelot
• PyPDF2
• pikepdf https://github.com/RichardScottOZ/pikepdf

OCR

• tesseract https://tesseract-ocr.github.io/tessdoc/Installation.html
• sudo apt install tesseract-ocr
• sudo apt install libtesseract-dev

Parsee pdf

• https://github.com/parsee-ai/parsee-pdf-reader
• pdf reader conflicts
• Installing collected packages: pytesseract, pypdf, pycparser, pdf2image, opencv-python, cffi, cryptography, pdfminer-six, parsee-pdf-reader Attempting uninstall: pypdf Found existing installation: pypdf 4.2.0 Uninstalling pypdf-4.2.0: Successfully uninstalled pypdf-4.2.0 ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. llama-index-readers-file 0.1.20 requires pypdf<5.0.0,>=4.0.1, but you have pypdf 3.17.4 which is incompatible. Successfully installed cffi-1.16.0 cryptography-42.0.7 opencv-python-4.9.0.80 parsee-pdf-reader-0.1.5.8 pdf2image-1.17.0 pdfminer-six-20221105 pycparser-2.22 pypdf-3.17.4 pytesseract-0.3.10

Metadata

• could use isfdb database dumps
• or rpggeek similarly

Logging

• needs logging added to track things in main program

Speed

• Game collection with lots of images - some things being old scans will be slow
• Consider parallelising
  • All at once embarassingly parallel job
  • Perhaps Lithops to abstract some configuration [supposedly]?
    • https://github.com/lithops-cloud/lithops

Lithops

Configure

• Config File
• Compute Backend [policy, role] https://github.com/lithops-cloud/lithops/blob/master/docs/source/compute_config/aws_lambda.md
• Storage Backend [bucket] - https://github.com/lithops-cloud/lithops/blob/master/docs/source/storage_config/aws_s3.md

Cloud test

/book-mentat/src/games$ lithops hello
2024-05-25 12:37:06,919 [INFO] config.py:139 -- Lithops v3.3.0 - Python3.10
2024-05-25 12:37:07,708 [INFO] aws_s3.py:59 -- S3 client created - Region: us-west-2
2024-05-25 12:37:09,791 [INFO] aws_lambda.py:97 -- AWS Lambda client created - Region: us-west-2
2024-05-25 12:37:09,793 [INFO] invokers.py:107 -- ExecutorID 8xc5f5-0 | JobID A000 - Selected Runtime: default-runtime-v310 - 256MB
2024-05-25 12:37:10,010 [INFO] invokers.py:115 -- Runtime default-runtime-v310 with 256MB is not yet deployed
2024-05-25 12:37:10,010 [INFO] aws_lambda.py:388 -- Deploying runtime: default-runtime-v310 - Memory: 256 - Timeout: 180
2024-05-25 12:37:10,838 [INFO] aws_lambda.py:187 -- Creating lambda layer for runtime default-runtime-v310
2024-05-25 12:38:53,088 [INFO] invokers.py:174 -- ExecutorID 8xc5f5-0 | JobID A000 - Starting function invocation: hello() - Total: 1 activations
2024-05-25 12:38:53,166 [INFO] invokers.py:213 -- ExecutorID 8xc5f5-0 | JobID A000 - View execution logs at /tmp/lithops-richard/logs/8dc5f5-0-A000.log
2024-05-25 12:38:53,191 [INFO] executors.py:491 -- ExecutorID 8xc5f5-0 - Getting results from 1 function activations
2024-05-25 12:38:53,191 [INFO] wait.py:101 -- ExecutorID 8xc5f5-0 - Waiting for 1 function activations to complete

Runtimes

• Need to build a Docker image for anything bespoke of interest

• Follow examples

  • https://github.com/lithops-cloud/lithops/tree/master/runtime/aws_lambda

• Needs Docker Desktop https://docs.docker.com/desktop/install/ubuntu/

At the end of the installation process, apt displays an error due to installing a downloaded package. You can ignore this error message.

N: Download is performed unsandboxed as root, as file '/home/user/Downloads/doc

• it has two timeouts - a runtime_timeout
• execution_timeout - in lithops section

Build

• lithops runtime build -f MyDockerfile -b aws_lambda my-container-runtime-name

• runtime memory –memory, -m Memory size in MBs to assign to the runtime.

• permissions issues:

  • https://stackoverflow.com/questions/51342810/how-to-fix-dial-unix-var-run-docker-sock-connect-permission-denied-when-gro

• list runtimes lithops runtime list -b aws_lambda

• libgl error

  • libgl1 library
  • suggestions https://stackoverflow.com/questions/55313610/importerror-libgl-so-1-cannot-open-shared-object-file-no-such-file-or-directo

• so good questions?

• other solutions

• pip install opencv-contrib-python

• install opencv-contrib-python rather than opencv-python.

• https://stackoverflow.com/questions/55313610/importerror-libgl-so-1-cannot-open-shared-object-file-no-such-file-or-directo

• At one point, I used opencv-python-headless which worked for my case with FastAPI when I deployed on Heroku once. What's the

• pip install -U opencv-python

• apt update && apt install -y libsm6 libxext6 ffmpeg libfontconfig1 libxrender1 libg

• runtime memory you can adjust in .lithops_config for pool

• /tmp is writeable in container

• Batch needs a service role that can do ECS

• also a job Role lithops-cloud/lithops#1359

• Choose Elastic Container Service on the use case list and then click on Elastic Container Service Task.

  • Click Next: Permissions. Select the policy created before (lithops-policy).
  • Click Next: Tags and Next: Review.
  • Type a role name, for example ecsTaskJobRole. Click on Create Role.

possibly

• Set the LD_LIBRARY_PATH environment variable ENV LD_LIBRARY_PATH="/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH"

Docker

• docker system prune --all --force

Failures

• lithops-cloud/lithops#1248