| name | knowledge-systems |
| description | Design knowledge retrieval and memory systems for AI applications. Covers wiki vs vector database tradeoffs (structured graph vs semantic similarity, when to use each, using both in one stack), LLM memory and retrieval (why LLMs forget between sessions, three approaches to persistence, retrieval vs behavior knobs), and knowledge system architecture (full pipeline from sources to rules to retrieval to output, three architectures compared with cost breakdown). Use when building RAG pipelines, choosing between vector databases and structured knowledge bases, designing memory persistence for agents, or architecting knowledge systems for AI products. |
| metadata | {"author":"Katrina Laszlo","version":"1.0.0"} |
Knowledge Systems
How to store, retrieve, and persist knowledge for AI systems. Three topics.
Wiki vs Vector Database
Two retrieval mechanisms solving the same problem differently.
Wiki: human pre-digests information into structured pages with explicit links. Total auditability. Scales to hundreds of pages with curation effort. Best for: research, strategy, competitive intel, onboarding.
Vector DB: math pre-filters by semantic similarity. Text converted to embeddings, nearby vectors = similar meaning. Scales to millions of chunks automatically. Best for: RAG, chat-with-docs, needle-in-haystack, fuzzy matching.
When to use both: wiki for curated context (entities, synthesis), vector for search (retrieval, similarity). Different layers of the same stack.
LLM Memory & Retrieval
LLMs forget between sessions. Three approaches to persistence: context injection (stuff relevant info into prompt), retrieval-augmented generation (RAG, search for relevant chunks), and fine-tuning (bake knowledge into weights).
Two knobs: retrieval (what information reaches the model) and behavior (how the model acts on it). Most problems are retrieval problems, not behavior problems.
Knowledge Architecture
Full pipeline: sources (documents, APIs, databases) to rules (CLAUDE.md, system prompts) to retrieval (vector search, wiki lookup) to output (generation, actions). Three architectures compared: monolithic context, RAG pipeline, agent with tools.
References
See references/ for detailed content on each topic.
