An agent plans the search and uses tools. Powerful, but more resource-intensive.
- Status
- Planned
- Priority
- 3 of 6
- Best use
- Questions that need two sources, or an assistant across many departments
- Accuracy and completeness
- Potentially high, but less predictable
- Cost and performance
- Higher cost and latency
Pick it when one question needs more than one source
Best for
- Questions that need two sources (a policy and a figure)
- Enterprise-wide assistants where people do not know which system holds the answer
- Two-part questions, and questions that need a calculation after retrieval
Ask it things like
- Which carrier has the highest total charges, and what is the daily pass price for Zone 2?
- How many lines are suspended, and who approved service order SO-88213?
- We had a 5-hour and a 30-hour outage this month. How many days of credit do we get?
Not the right choice for
- High-volume, simple questions where speed and cost matter most (use Vector or Hybrid)
- Anything that must give the same path every time for an audit, unless the plan is logged (it is in the Lab)
Choose it over Hybrid when questions regularly mix documents and tables, or users ask across departments. On the Altamira scorecard Agentic answered the most questions of any method.
Choose it over GraphRAG when the answer comes from a few sources, not from following a chain of relationships.
- Data it reads
- Everything the other methods read; it calls them as tools.
- How it gets the data
- Reads the question's signals (figure words, matching tables, exact codes), picks a tool, checks the evidence, retries with a different tool once, and declines if nothing is supported.
- How fresh the answer is
- Same as the tool it calls: live for spreadsheets and source systems, as of the last index for documents.
- Relative cost
- Moderate to high
How Agentic RAG works
- 1
Plan
Break the request into bounded retrieval subtasks and required evidence types.
- 2
Select
Choose applicable Hybrid, Structured, Graph, Multimodal, Memory or other approved customer tools.
- 3
Retrieve
Execute each tool under the effective user and tenant authorization; preserve source references, tool outputs and traces.
- 4
Evaluate & Iterate
Assess relevance and evidence sufficiency. Retry or switch tools only within iteration, latency and cost budgets; recognize non-answerable requests.
- 5
Draft
Prepare a response only from authorized evidence and distinguish source facts from inferred synthesis.
- 6
Validate & Return
Check claims, citations and policy requirements; withhold unsupported claims, escalate where required, or explicitly abstain.
Use it only when a question genuinely needs several retrieval steps or tools.
One router picks the right method for each question
Context66 combines these methods as Ontology-guided Adaptive RAG. Every answer is checked against its evidence, and the system abstains when the evidence is incomplete.
- User question
- Intent and context detection
- Ontology and semantic interpretation
- Query router
All branches, orchestrated by the agent
- Evidence fusion and ranking
- Answer written only from retrieved evidence
- Claim-level validation and citation checking

