From archive to graph

Pipeline

1. 01

   Digitization

   High-resolution capture with color and scale references. Vision models assess each page's physical condition and produce restoration reports.

   • High-resolution capture
   • Color charts
   • Vision LLMs
2. 02

   Assisted cleanup

   Language models correct OCR errors and normalize transcriptions while preserving historical spelling where it carries meaning.

   • Llama 2 7B
   • Format normalization
3. 03

   Entity recognition

   Automated extraction of people, places, institutions, and events, with temporal alignment and semantic embeddings across the corpus.

   • Flair NER
   • Semantic embeddings
   • Temporal alignment
4. 04

   Human validation

   Every annotation goes through expert review in our own tooling and in Argilla. No entity reaches the graph without human and vocabulary control.

   • Argilla
   • In-house NER corrector
   • Controlled vocabularies
5. 05

   Network modeling

   Concurrence graphs, community detection, and comparison against synthetic realizations to infer social dynamics from partial observations.

   • Concurrence graphs
   • Leiden / Louvain
   • Barabási–Albert

Stack