Open-source RAG, vector search and agent memory

Questions people ask

Do I still need RAG with a million-token context window?

Usually yes. Long context solves reading; it does not solve cost, latency or knowing which documents to send. Retrieval remains the cheaper answer whenever the corpus grows faster than the budget.

Which part of a RAG pipeline should I build myself?

The chunking and the evaluation, because both are specific to your documents. Storage and embedding are solved and worth taking off the shelf.

How do I know retrieval is working?

Build a small set of questions with known correct sources before you build the pipeline, and measure whether the right source comes back. Without that set you are judging fluency, not accuracy.