The Journal

My grandfather served as a soldier in the US Army's 48th Battalion in WWII from 1944 to 1945, a period when the Allied military crossed the German military's last major line of defenses and consolidated territorial gains leading up to the announcement VE Day on May 8th. A few years after my grandfather passed away, my family was looking at his handwritten journal, a detailed description of his experiences he carried through France and into Germany from 1944 to 1945. I took photos of the all the pages to make sure we preserved the text. I also decided to transcribe the text into a Google doc and share with my family since physically handling the journal would only accelerate its deterioration.

After completing the transcription, I noticed how abruptly the story ended. It was March 1945 and he was on his way back to his platoon in Lupstein, France after recuperating from an injury.

I'm not sure why he stopped. My grandmother said after he returned, his group was moving too fast for him to spend time writing. It's quite possible he thought he would pick up after the war. But while I wanted to preserve the integrity his original text, leaving the story without a conclusion was unsatisfying.

WE CAME TO FIGHT

After the war ended and he returned home, the Army sent him a book, “We Came to Fight! - History from New York to Jettenbach”, recounting the 48th Battalion's experiences from training at Camp Campbell, KY to fighting in France and Germany.

I also transcribed the full text from this book since, like my grandfather’s journal, it was not in good shape. Toward the end of the transcription, I realized I could just summarize the Battalion's activities from March to May as a conclusion to the journal. The problem was, the full text of “We Came to Fight” was greater than 11,500 words and I was too worn out from all the transcribing to attempt a summarization myself.

LLAMA-INDEX

Around January 2023, I discovered GPTIndex, an early framework for connecting documents to large language models. The project was later renamed LlamaIndex, and it quickly became a useful tool for building applications that let LLMs reason over external text. It seemed like a perfect fit for generating the historical summary I needed.

I used GPTIndex with OpenAI’s text-davinci-003 model to produce an initial summary of my grandfather’s final months in World War II. Drawing from the documents he left behind and other historical sources, the model helped assemble a plausible timeline of his journey out of Germany and back to the United States at the end of the war. Later, when revisiting the project, I reran the workflow using a newer version of LlamaIndex and GPT-4, which generated a more detailed account.

At the time, I simply didn’t have the time or emotional bandwidth to manually reconstruct the final chapter of my grandfather’s story. If you follow AI development, you’ve probably seen far more sophisticated applications of LLMs. In comparison, document summarization might seem basic. But in this case, it had real personal meaning. Even a relatively simple capability, summarizing thousands of words into a coherent narrative, helped me complete a project that had sat unfinished for decades.

Below is the original GPT-3 summary. The notebook also includes the later GPT-4 version for comparison.

THE CONCLUSION