Authoritative answers from governed enterprise data and queries.
- Status
- Planned
- Priority
- 2 of 6
- Best use
- Totals, counts, rankings and exact records from governed tables, queried live
- Accuracy and completeness
- High factual precision for governed structured data
- Cost and performance
- Low to moderate runtime cost
Pick it whenever the answer is a number from a table
Best for
- Totals, counts, averages, rankings and trends over tables
- Looking up one record exactly (one invoice, one line, one employee ID)
- Figures leadership will act on, where a wrong number is expensive
Ask it things like
- Which carrier has the highest total charges?
- How many lines are suspended?
- What were the total zero-usage charges for AT&T in March 2026?
Not the right choice for
- Anything written in prose: policies, contracts, emails (use Hybrid)
- Scans and pictures of tables (use Multimodal first, then query the extracted table)
Choose it over Hybrid when the answer is a number computed from many rows. Hybrid can only quote rows (it hands totals to this method), and quoting a few rows for a total is how wrong answers happen.
Choose it over Agentic when the question is clearly about one table. Agentic would route here anyway, after an extra planning step.
- Data it reads
- Spreadsheets (CSV, Excel) loaded as tables, or governed tables and views in a warehouse through the semantic layer.
- How it gets the data
- Plans a SQL query from the question and the table's columns, checks it with a SQL gate (read only, row cap, no other tables), runs it, and checks every figure in the answer against the result.
- How fresh the answer is
- Live: the query runs at question time, so a changed row shows up in the next answer. This is the method to show real-time retrieval from the source.
- Relative cost
- Lowest
How Structured RAG + semantic layer works
- 1
Discover Approved Sources
Load permitted schemas, API contracts, keys, examples, data ownership and available metric definitions; do not assume cataloged means approved.
- 2
Interpret and Resolve Terms
Map user terms and metrics to validated schema concepts or approved semantic definitions; clarify unresolved material ambiguity.
- 3
Plan and Authorize Query
Build SQL/API requests using approved metadata, templates or a semantic query service; validate tables, joins, aggregates, parameters, cost and row/column policy before execution.
- 4
Execute at Source
Use least-privilege, bounded, read-only access and execute calculations in source systems or a governed metric engine.
- 5
Return Traceable Results
Preserve returned values with source identifiers, query/filters, units, data freshness and as-of context where available.
- 6
Draft Explanation
Explain the returned structured evidence in plain language without substituting LLM-generated metric definitions or arithmetic.
- 7
Verify and Return
Compare stated values and units with the query result; enforce output security and abstain on unresolved ambiguity or failed validation.
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
This method
- Evidence fusion and ranking
- Answer written only from retrieved evidence
- Claim-level validation and citation checking

