Loading patterns...
If this never updates, open the site over HTTP(s) so data can load.
graphslop
A field guide to LLM graph extraction gone wrong — and the occasional redemption arc.
patterns
Common mistakes that show up in production pipelines. If you recognize your setup, we're not judging — we're judging the graph.
slop of the day
One highlighted example, rotated by calendar day (UTC).
Loading...
archive
Examples of extraction gone wrong.
Loading...
what works
Nothing glamorous, which is how you know it's real engineering.
Structure first, triples second
Parse documents into logical units (sections, tables, lists) before asking an LLM to assert relations. Text locality matters — fake locality produces fake edges.
Entity resolution is not optional
Canonical IDs, blocking, pairwise scoring, and human review for long tails. Lowercasing strings is not a merge strategy.
Schema as a contract
Version your ontology, constrain predicates, and test extraction against gold sets. If the model drifts types every deploy, you don't have a graph.
Evaluation that matters
Edge-level precision/recall, constraint violations, and retrieval quality beats "the graph looks big." Big graphs full of slop are just high-entropy compost.
Provenance everywhere
Store source spans, document IDs, and extraction version. When someone asks why an edge exists, the answer should not be vibes.
Hybrid pipelines
Rules and dictionaries for stable relations; LLMs for fuzzy bits — with calibration and abstention when confidence is low.