My starting hypothesis: RAG is an evidence-access design problem, not a vector-database choice. Compare whole-document reading, hybrid retrieval and agent-directed search on actual questions; check product maturity separately from the architecture's name.

Initial RAG landscape: evidence access, not a single winning stack

Initial source review for the RAG investigation, checked September 29, 2026. This is a starting interpretation, not a completed benchmark or an exhaustive product ranking.

1. Long context and retrieval are not mutually exclusive

The 2025 LaRA study found that the relative performance of retrieval-augmented generation and long-context reading depends on the model, task, input length, and retrieved material. Its results do not establish a universal winner, and should not be treated as measurements of September 2026 models.

My implication for this investigation: include whole-document reading as a serious baseline, and test combinations that retrieve a document before reading it in detail. Do not assume either that retrieval is always necessary or that a large context window removes evidence-selection decisions.

2. A stronger baseline than vector similarity alone

Anthropic's Contextual Retrieval experiment, published September 19, 2024, combines document-aware chunk context, semantic retrieval, BM25, and reranking. In its tests, the combined approach reduced top-20 retrieval failures from 5.7% to 1.9%. That is a retrieval metric in a particular experiment, not a 67% improvement in final answer correctness.

I would use hybrid lexical and semantic retrieval with reranking as a candidate baseline, then test whether contextual enrichment pays for its preprocessing cost on the target corpus.

3. Agents can search instead of accepting one fixed context bundle

Anthropic's context-engineering account, published September 29, 2025, describes combining upfront context with just-in-time discovery. Its Claude Code example uses file navigation and search primitives rather than requiring every retrieval action to pass through a vector index.

This supports a distinction between retrieval as a fixed pipeline and retrieval as a tool used during an agent's work. It does not prove that agent-directed search is faster or more accurate for every workload. A parallel multi-query service and a genuinely iterative search agent should also be evaluated separately.

4. Product status must be checked feature by feature

Gemini API File Search offers managed import, chunking, indexing, and retrieval. Its current documentation includes text and image/multimodal embeddings, while explicitly excluding audio and video formats. Managed RAG is therefore a product category, not merely a collection of framework recipes.

Azure AI Search documentation, updated September 17, 2026, distinguishes generally available minimal extractive retrieval with supported knowledge sources through API 2026-04-01 from preview capabilities through 2026-08-01-preview. LLM query planning, answer synthesis, and non-minimal reasoning remain preview. The label "agentic retrieval" alone does not establish production readiness for every feature.

The Microsoft GraphRAG repository currently says it is largely in maintenance mode, with no new features or incoming PRs, while retaining appropriate fixes and dependency updates. This is a statement about that implementation, not evidence that graph-based retrieval as a class has stopped developing. Its documented architecture uses entities, relationships, and community summaries for local and corpus-wide questions.

5. Correct retrieval is only one part of reliability

The Sufficient Context study distinguishes missing evidence from failures to use available evidence, and examines answering versus abstaining. For our evaluation, finding relevant material, having enough evidence, producing a correct answer, and supporting claims with citations should be separate checks.

Permissions also need operational verification. Azure's access-control documentation explains that indexed authorization metadata must be synchronized: source permission changes are not necessarily reflected immediately. An access-control feature is not, by itself, proof of instant revocation across an indexed corpus.

First practical question

What changes when the same representative questions are answered using whole-document reading, hybrid retrieval with reranking, and agent-directed search? Compare correctness and evidence support alongside latency, cost, freshness, and permission behavior. Add graph-based retrieval or a persistent memory layer only when a diagnosed requirement justifies the extra architecture.

I propose starting with that comparison before choosing a vector database or building a broad vendor shortlist.