ModelRefs / Agentic RAG — Canonical Workflow
Agentic RAG — Canonical Workflow
Agentic RAG: provisional AI workflow implementation reference with candidate models, providers, tools, and architecture.
Overview
Agentic RAG upgrades a standard single-shot retrieval pipeline with a planning agent that can reformulate queries, switch retrieval strategies, validate the retrieved evidence against the question and iterate until the answer meets a confidence threshold. The agent maintains a scratchpad of retrieval attempts, surfaces the chain of retrieval decisions alongside the final answer, and falls back gracefully when no confident answer can be constructed. This pattern is recommended over single-shot RAG for complex multi-hop questions and research-grade applications where source traceability is critical.
Implementation profile
| Category | agentic-models |
|---|---|
| Implementation maturity | production |
| Evidence status | partial |
| Primary use cases | rag, agents, reasoning |
| Deployment options | managed-api, hybrid |
| Architectures | serverless-api, managed-container, self-hosted-cluster |
Candidate models with published references
- BGE-M3
- GPT-5
- GPT-5 Mini
- Claude Opus 4
- Llama 4 Scout
- DeepSeek R1
- Mistral Large 2
- Command R+
- o3
- o4 Mini
- Text Embedding 3 Large
- Claude Sonnet 4
Coverage means the model is a candidate worth evaluating for this workflow, not a ranking or a recommendation. Models whose reference pages are still in review are omitted.
Benchmarks relevant to this workflow
miracl, mkqa, mldr, swe-bench, aider-polyglot, gpqa, aime-2025, tau-bench, browsecomp-long-context, longfact-concepts, terminal-bench, mmmu, mmlu-pro, livecodebench.
Relevance is a coverage signal from the canonical registry. Each benchmark only describes its own protocol and date, so confirm the harness matches your workload before treating a score as evidence.
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Agentic RAG — Canonical Workflow.