ComponentsBuild with this
rag-wikiemerging
LLM Wiki
Karpathy-style compounding knowledge base. Gets smarter over time.
Overview
Inspired by Andrej Karpathy's approach — the LLM continuously reads, synthesizes, and writes knowledge articles from your corpus into a structured wiki. Over time the wiki becomes dense with cross-referenced insight, and retrieval operates over high-quality synthesized knowledge rather than raw chunks.
Architecture
Interactive walkthrough01
Ingestion
Document Loadercorpus intake
Topic DiscoveryLLM clustering
Article Writerone wiki page / topic
Article Storeembedded summaries
02
Retrieval
Query Embeddingsame model
Cosine Similarityarticle ranking
Top Articleswhole-topic context
03
Generation
Synthesis Promptarticle context
LLM Calltopic-aware answer
01
Ingestion
Document Loadercorpus intake
Topic DiscoveryLLM clustering
Article Writerone wiki page / topic
Article Storeembedded summaries
02
Retrieval
Query Embeddingsame model
Cosine Similarityarticle ranking
Top Articleswhole-topic context
03
Generation
Synthesis Promptarticle context
LLM Calltopic-aware answer
Summarized pipeline view. For the full interactive, scroll-driven walkthrough with clickable stages → Pipeline detail
When to use
Use when
- corpus_size < 1000 documents
- corpus is static or slow-changing
- queries are conceptual / analytical
- knowledge accumulation is the goal
Avoid when
- corpus updates frequently
- queries are time-sensitive
- cost constraints are tight
- real-time retrieval needed
Compatible vector databases
ChromapgvectorQdrant
Compatible frameworks
raw pythonlangchain
#wiki#karpathy#knowledge-synthesis#compounding#static-corpus
Ready to build with LLM Wiki?
Walk through the wizard to generate a complete, parameterized pipeline.