How I built a peptide knowledge base using the Karpathy method.

We took 15 biomedical sources - academic papers, market reports, regulatory documents - and turned them into a living, interlinked knowledge base using a four-stage process. Here's how it works at a high level.

01 - Extract

We studied the document archive and identified the knowledge structures buried inside: entities (companies, regulators, labs), products (peptides, therapeutics), markets (segments, pricing, access channels), and concepts (mechanisms, delivery methods, regulatory frameworks). Each source was read once, deeply, and the key information was pulled out - not summarized, but structurally decomposed into atomic fragments.

02 - Structure

We designed custom templates for each page type - product, concept, entity, market, source - so every fragment follows a consistent schema. A product page always has the same fields. A market page always has the same sections. This consistency is what makes the system compoundable: new information slots into a known structure instead of floating as unstructured text.

03 - Connect

Every fragment was cross-linked into an interlinked knowledge graph. A company mentioned in three reports gets one page with three sources. A peptide referenced across a clinical paper, a market report, and a regulatory notice gets all three perspectives unified on a single page. Contradictions between sources are flagged, not hidden. The graph view reveals clusters, gaps, and relationships that no single document could show.

04 - Compound

With the base built, every new source and every new question strengthens the whole. A query about "commercially in-demand peptides" draws on 50+ product pages, 5 market analyses, and 7 entity profiles to produce an answer no single search could replicate - and that answer gets filed back into the wiki, making the next query even richer. The knowledge compounds.