NVS Research · Retrieval

A RAG agent doesn't know the answer. It goes and gets it.

Retrieval-augmented generation gives a language model a second channel: instead of answering from memory, it pulls the relevant documents first and answers from those. The agentic version adds a loop that decides when and what to fetch. Both are real progress. Neither is magic, and the honest engineering matters more than the hype.

guess (high uncertainty) retrieved document = the missing bits grounded answer
Retrieval closes the gap between what the model knows and what the question needs.
// retrieval as a channel: the bits it adds to the answer I(A; C) = H(A) − H(A | C) H(A) the answer's uncertainty before retrieval H(A|C) what's left once the context C is in hand
01

Naive RAG retrieves once. An agent retrieves like it means it.

Naive (standard) RAG is one fixed pass: take the question, pull the top few matching chunks from a document store, staple them to the prompt, generate once. Simple and fast, and often enough.

Agentic RAG puts retrieval inside a decision loop. The model can decide whether to retrieve at all, rewrite a vague question into a better search, retrieve again across multiple hops, judge whether what came back is any good, and correct course before answering. That's the real definition. Not "RAG plus buzzwords," but retrieval placed under an agent's control (the framing in the 2025 Agentic RAG survey).

The catch, proven not asserted: that loop is not free. Measured against a well-tuned fixed pipeline, agentic RAG used roughly 3.3× the input tokens, 1.9× the output tokens, and ~1.5× the time. On one fact-checking benchmark it actually lost by 28.8 F1 points, because it kept retrieving when it should have just answered. Sophistication can make things worse.

3.3× tokens  ·  1.5× time
Agentic RAG vs. a well-optimized fixed pipeline (Ferrazzi et al., "Is Agentic RAG worth it?", 2026). Use the loop where multi-step reasoning earns its cost, not everywhere.
02

The winning design is boring: search two ways, then re-rank.

Across independent 2026 benchmarks, one pattern beats everything else by wide, statistically significant margins, and it isn't the clever agent. It's a two-stage pipeline: retrieve with both keyword search (BM25) and semantic embeddings, fuse the two lists, then pass the top results through a reranker that reads query and document together and re-orders them.

On 23,088 real financial-document questions, that hybrid-plus-reranker design hit Recall@5 of 0.816, versus 0.587 for semantic embeddings alone. Same documents, same questions; the difference is the pipeline.

Stage 1 · cast wide
Hybrid retrieval
Keyword (BM25) + semantic (embeddings), fused. Each catches what the other misses.
Stage 2 · sort hard
Reranker
A cross-encoder reads query + document together and re-orders by true relevance.
Result
The right 5
The chunks the answer actually needs, at the top, where the model will use them.
0.587dense only
0.644keyword only
0.695hybrid
0.816hybrid + rerank
Recall@5 on T2-RAGBench (23,088 financial QA queries, 2026). Higher = the right document was in the top 5 more often.
03

Semantic search doesn't always win. Sometimes plain keywords do.

The industry reflex is "embeddings understand meaning, so semantic search is strictly better." It isn't. Financial documents are full of exact terms: company names, ticker symbols, metric labels, and fiscal quarters. On those, classic keyword search (BM25) beat a state-of-the-art embedding model on 9 of 10 retrieval metrics. Precise language is a strong signal, and embeddings blur it.

The lesson isn't "keywords beat embeddings." It's that the right retrieval depends on your content, and the honest way to know is to measure, not to buy whichever technique is fashionable. Even the fashionable self-correcting agent (Corrective RAG) fired on 63% of queries in one study and still landed below a plain hybrid search. A clever loop can't rescue a first stage that was built on a guess.

04

RAG is for words, not numbers.

The single most useful thing to know before building anything: RAG is a tool for long-form, unstructured text. Emails, meeting transcripts, documents, knowledge bases, the prose a business generates. There, retrieval shines: it finds the relevant passage in a haystack of writing and hands it to the model.

It is the wrong tool for structured, tabular data. "How many orders shipped last month" is a question with an exact answer that lives in a database. You ask it with a query, not by embedding spreadsheet cells into a vector store and hoping the right rows float up. Forcing structured data through retrieval is how you get confident, wrong arithmetic.

Prose a person wrote or said
→ RAG / retrieval
Emails, transcripts, docs, policies, notes. Meaning-bearing text where the answer is "find the passage that says it."
retrieve · ground · cite
Records, metrics, tables
→ database + query
Orders, amounts, dates, statuses. Exact values where the answer is a calculation, not a paraphrase.
select · aggregate · compute
The test before you build: is this prose a person wrote, or is it data? Prose retrieves; data gets queried.
05

You can't improve what you don't measure, and most measures lie.

A RAG system that sounds fluent can still be quietly wrong. The way you catch that is groundedness: break the AI's answer into individual claims and check each one against the documents it retrieved. If a sentence isn't supported by the source, it's a hallucination, no matter how confident it reads. (This is what tools like RAGAS Faithfulness automate: supported claims ÷ total claims.)

The uncomfortable part: when researchers tested the popular automatic quality scores against human judgment, most of them correlated poorly, or even negatively. Only measuring at the individual-claim level reliably tracked whether an answer was actually grounded. The takeaway isn't "don't measure." It's that a real evaluation loop is the moat, the difference between a demo that impresses and a system you can trust in front of a customer.

// groundedness, measured claim by claim faithfulness = supported claims / total claims decompose the answer, check each claim against the context
The honesty rule

Where this is rigorous, and where it's still fragile.

A research page should invite scrutiny and survive it. So, plainly:

  • Rigorous: retrieval genuinely is an information channel. It injects the specific bits an answer was missing. That part is not analogy; it's the same information theory the rest of our research runs on.
  • Rigorous: the benchmark numbers here (hybrid + rerank, the cost of agentic loops, keyword-beats-embeddings on precise text) come from 2026 empirical studies with real datasets. But they're recent preprints on specific domains, so read them as strong current evidence, not settled law.
  • Still fragile: RAG agents are reliable when the retrieval is well-built and evaluated, and unreliable when it isn't. Self-correction helps but doesn't rescue a weak foundation, and long context windows don't make retrieval obsolete. Stuffing everything into the prompt scatters the model's attention, and the cost grows faster than the text.
  • Be skeptical of round numbers: vendor claims like "+42% faithfulness" tend to arrive with no dataset and no method behind them. If a figure can't tell you how it was measured, treat it as marketing.
+39%
Recall@5 gain of hybrid retrieval + reranking over semantic embeddings alone (T2-RAGBench, 23,088 queries).
9 / 10
Retrieval metrics on which plain keyword search (BM25) beat a state-of-the-art embedding model on financial documents.
3.3×
Input tokens an agentic RAG loop spent vs. a well-tuned fixed pipeline, at times while underperforming it.
claim-level
The only automatic grounding measure that reliably tracked human judgment. Whole-answer scores didn't.
Sources. Singh et al., "Agentic RAG: A Survey" (arXiv:2501.09136) · Ferrazzi et al., "Is Agentic RAG worth it?" (arXiv:2601.07711) · "T2-RAGBench" (arXiv:2604.01733) · SemEval-2026 multi-turn RAG (arXiv:2606.28352) · "Seeing Through the MiRAGE," metric reliability (arXiv:2510.24870) · RAGAS (Es et al., 2024). Figures attributed to their studies; recent preprints, not settled law.
The NVS take

Retrieval isn't magic. It's engineering you can measure.

A RAG agent is only as good as what it retrieves and how honestly you test it. We build the boring parts right, the retrieval that fits your content and the evaluation loop that proves it, then reserve the clever agent for where it earns its cost.
Talk to us about retrieval done right →