Choosing retrieval.
For the question.

Retrieval-augmented generation brings external evidence into a model’s working context. Vector search is one implementation; an agent can also search keywords, inspect document structure and read relevant passages through tools. These approaches can be combined. [01] [02]

For my curated Markdown notes, I treat agentic file search as a useful baseline. It lets an agent refine a query and inspect the original text without maintaining an embedding index. Subramanian and colleagues found that tool-based keyword search approached conventional RAG performance on the evaluated tasks. That supports testing a simpler approach, with results checked on the actual collection. [01]

I would compare approaches against representative questions: exact lookups, terminology mismatches, questions spanning several sources and questions that require a current structured record.

A relevant passage can exist in the source collection and still fail to reach the model. This is a retrieval failure. It differs from an answer misusing evidence that was successfully retrieved.

My first checks would be the query, document boundaries, missing surrounding context and ranking. Hybrid retrieval combines lexical and semantic matching; reranking can help select relevant candidates. Anthropic’s Contextual Retrieval work also tests adding document-specific context to chunks before indexing. Its reported improvements belong to the evaluated setup. [02]

I want to measure both evidence retrieval and the resulting answer. More passages help when they improve correctness, completeness or traceability. Latency and maintenance matter too, especially for a voice interface.

GraphRAG becomes interesting when questions depend on relationships or broader themes across documents. Microsoft’s original approach builds an entity graph and community summaries to support questions about the collection as a whole. [03]

GraphRAG-Bench offers a useful qualification: graph methods showed advantages on some complex reasoning and summarization tasks, while basic RAG remained competitive on simple fact retrieval. [04]

My interpretation is to introduce a graph when representative questions demonstrate that relationships are the missing capability. The graph also needs to stay aligned with its sources. For frequently edited notes, I would first test clearer structure, better retrieval and an iterative search loop.