Quick brush-up. "GenAI exposure" is a plus on the JD — enough to speak credibly, tied to your NLP background.
Foundations
What is an LLM / how does it work (gist)?
A large neural network (transformer) trained to predict the next token. Given a prompt, it generates text token-by-token. Not a database — it generates, which is why it can hallucinate.
What is a token / embedding?
Token = a chunk of text (~¾ of a word). Embedding = a vector (list of numbers) capturing meaning, so similar text has nearby vectors. Basis of semantic search.
What is RAG (Retrieval-Augmented Generation)?
Instead of relying only on the model’s training, you retrieve relevant documents and feed them into the prompt as context. Flow: chunk docs → embed → store in a vector DB → at query time, embed the question, find similar chunks, stuff them into the prompt → LLM answers grounded in them. Fixes hallucination + stale knowledge.
Building blocks
Vector database?
Stores embeddings and does similarity search (nearest-neighbor). E.g. pgvector, Pinecone, Chroma, Elasticsearch.
LangChain?
A framework to build LLM apps — chains (compose steps), tools (let the LLM call functions/APIs), memory (conversation history), retrievers for RAG. Glue for LLM pipelines.
LangGraph?
Builds stateful, multi-step agent workflows as a graph (nodes = steps, edges = transitions, with state + checkpoints). For agents that loop, branch, and maintain state — beyond a linear chain.
Spring AI (YOUR lane — Java + AI)
Do RAG in Java/Spring:
VectorStoreabstraction,ChatClientfluent API,QuestionAnswerAdvisor(out-of-box RAG) orRetrievalAugmentationAdvisor(modular). Lets you build AI on your existing stack. This is your differentiator — Java dev who ships AI.
Positioning
Your honest AI story
“I have a data-science masters and did classical NLP with OpenNLP and Solr. Now I’m building on that with GenAI — I understand RAG end-to-end (chunk → embed → vector store → retrieve → generate), and I’m learning LangChain, LangGraph, and Spring AI so I can do it in Java.” Credible, tied to real background.
Drill cards
What is RAG?::Retrieve relevant docs (via embeddings + vector search), inject them into the prompt as context so the LLM answers grounded in real data. Fixes hallucination + stale knowledge.
What is an embedding?::A vector representing text meaning; similar text → nearby vectors. Enables semantic/similarity search.
LangChain vs LangGraph?::LangChain = compose LLM steps/tools/memory (often linear). LangGraph = stateful multi-step agent workflows as a graph (nodes/edges/checkpoints).
What is a vector database?::Stores embeddings and does nearest-neighbor similarity search (pgvector, Pinecone, Chroma).
Why do LLMs hallucinate?::They generate the most likely next token, not retrieve facts — so they can produce plausible-but-wrong output. RAG grounds them in real docs.
Spring AI?::Build RAG/LLM apps in Java — VectorStore, ChatClient, QuestionAnswerAdvisor. AI on your existing Spring stack.