While some proclaim that "RAG is dead," giving LLMs access to internal knowledge remains one of the primary ways to leverage AI in the enterprise. What has changed since the original RAG paper was published in 2020 are the strategies systems use to select data for answering questions.

Many modern approaches are agentic: instead of searching for relevant documents once, agents can repeatedly search and filter before answering. They can also query databases directly using languages such as SQL or Cypher.

Benchmark and systems

We use PlaceboBench, a dataset of 69 questions from real clinicians about prescription medicines. The questions are answered using product information published by the European Medicines Agency (EMA).

The four systems tested were:

All four systems use GPT-5.4 with thinking set to "high." It has the lowest hallucination rate on PlaceboBench and has performed well on this task.

Evaluation

We report two metrics: hallucination rate and pairwise win rate. Hallucination rate measures the percentage of answers containing claims unsupported by the retrieved evidence. Win rate compares answers from two systems and asks an LLM judge which is better.

We measure hallucination rate with a Blue Guardrails evaluator that checks whether claims are grounded in the retrieved evidence. The evaluator achieves an F1 score of 0.85 against human review on PlaceboBench.

For win rate, an LLM judge receives the question, two system answers, and the detected issues in each answer. It considers hallucination severity, relevance, clarity, and depth before deciding whether System A wins, System B wins, or the systems tie.

We shuffle the system order to reduce presentation bias and use Gemini 3.8 Flash as the judge. Each sample is judged five times, and we report the average win rate.

Results

Hallucination rate

Confirmed hallucination counts and rates by RAG system

System	Answers with confirmed hallucinations	Rate
Hybrid	6 / 69	8.7%
Graph	10 / 69	14.5%
Agentic	21 / 69	30.4%
Naive	16 / 69	23.2%

Hybrid RAG has the lowest hallucination rate at 8.7%, followed by agentic graph RAG at 14.5% and naive RAG at 23.2%. Agentic vector RAG has the highest rate at 30.4%.

At 8.7%, hybrid RAG would rank among the best-performing production systems evaluated by Blue Guardrails.

Pairwise win rate

HybridGraphAgenticNaive

Hybrid winsTieHybrid losesError bars: ±1 SD

Pairwise judge results with mean and standard deviation in percentage points

Comparison	Win	Tie	Loss
Hybrid vs. Graph	49.0% ± 1.6 pp	31.6% ± 1.6 pp	19.4% ± 1.3 pp
Hybrid vs. Agentic	51.6% ± 2.2 pp	25.2% ± 1.9 pp	23.2% ± 2.3 pp
Hybrid vs. Naive	61.4% ± 1.7 pp	10.7% ± 1.9 pp	27.8% ± 3.5 pp
Graph vs. Agentic	39.7% ± 1.3 pp	30.1% ± 1.9 pp	30.1% ± 1.2 pp
Graph vs. Naive	53.0% ± 1.3 pp	6.7% ± 1.7 pp	40.3% ± 1.9 pp
Agentic vs. Naive	51.6% ± 2.2 pp	6.1% ± 2.8 pp	42.3% ± 1.2 pp

The trend continues in the pairwise comparisons. Hybrid RAG produced better answers than all other systems in this evaluation. Agentic graph RAG wins two of its three pairings, although its advantage over agentic vector RAG is narrow. Naive RAG produces the lowest answer quality on average.

Agentic vector RAG has the highest hallucination rate, while naive RAG ranks lowest in pairwise answer quality. Manual inspection helps explain the difference: hallucination rate counts every affected answer equally, while the pairwise judge also considers hallucination severity, relevance, clarity, and depth.

Operational metrics

Quality alone is not enough to choose the right system. We also compare cost per task, duration, and tool-call statistics across the four approaches.

Generation latency per task

Bar chart of average summed model-generation duration per task for four RAG systems

System	Generation duration
Hybrid	44.6s
Graph	41.9s
Agentic	27.3s
Naive	11.2s

We consider model-request duration only, excluding tool execution time because it depends heavily on the technology, infrastructure, and dataset size.

Hybrid RAG and agentic graph RAG take more than 40 seconds on average. Agentic vector RAG takes just over 27 seconds, while naive RAG is fastest at 11.2 seconds.

Cost per task

Bar chart of average generation cost per task for four RAG systems

System	Cost per task
Hybrid	$0.177
Graph	$0.148
Agentic	$0.140
Naive	$0.045

The higher quality of hybrid RAG comes at a cost: average inference spend is $0.177 per task. Agentic graph and agentic vector RAG cost between $0.14 and $0.15, while naive RAG costs roughly one-quarter as much as hybrid RAG.

Tool calls per task

Bar chart of average tool calls per task for four RAG systems

System	Average tool calls
Hybrid	9.5
Graph	7.3
Agentic	9.5
Naive	0.0

Agentic graph RAG makes 7.3 tool calls per question, compared with roughly 9.5 for hybrid and agentic vector RAG. Naive RAG is not agentic and therefore makes no tool calls.

Conclusion

The right approach depends on the relative importance of answer quality, latency, and cost. Hybrid RAG is the strongest option when answer quality is the primary objective. Agentic graph RAG is the closest alternative, averaging $0.148 rather than $0.177 per task and 41.9 rather than 44.6 seconds. Naive RAG is best suited to latency- and cost-sensitive applications, averaging 11.2 seconds and $0.045 per task. Agentic vector RAG was faster than the graph and hybrid systems at 27.3 seconds, but had the highest hallucination rate at 30.4%.