What to check before a query runs
Confirm that the tables and columns exist, the joins and aggregations mean what the business expects, and the requester has access.
Knowledge Core · AI Ontology
Turn scattered data
into knowledge AI can use
ONDO AI ONTOLOGY connects documents, data structures, and business terms in a knowledge graph. It helps conversational AI answer questions using enterprise knowledge.
Connect internal knowledge · validate queries before execution · apply knowledge changes after human approval
- ConnectOrganize knowledgeDocuments, schemas, business terms
- CheckValidate SQL firstRegistered tables, columns, joins
- ReviewReview the evidence behind each queryRejection reasons and business meaning
- ApproveImprove knowledgeReview changes and retain history
The Problem
AI needs to understand your company's data rules
AI-generated SQL can refer to tables or columns that do not exist. Connect internal data structures with business terms and check the query before it runs.
Validation Design
Generate queries with internal context,
then validate before execution
The graph supplies relevant structures and business terms for SQL generation. A separate validation step checks the result against registered structures and rules.
Pre-execution validation · ONDO AI ONTOLOGY
Compare generated SQL with the registered graph and validation rules. Reject violations and record why, so queries and knowledge can be improved.
- Schema checks
- Separate generation and validation
- Rules before execution
- Rejection history
Ask your data and check the evidence
Ask a question in natural language to find the data you need and see it in a table. Review the query conditions and results in one place.
This example shows a product query in the ONDO AI workspace. Products and figures shown are sample data.
Grounding on — answer backed by verified evidence
MCP integrationQuery data from internal databases and tools
Deep researchMulti-step, in-depth research
General searchQuick factual search
How can I help with your work today?
Brand: NORDÉ Year: 2026 Season: SS Item: TS
Show the 10 products with the lowest cost ratios under these conditions, ordered by sell-through rate, along with their supplier information.
ONDO AI
The results are ready. Here is the summary.
📊 NORDÉ 2026 SS / TS items — 10 lowest-cost-ratio products (by sell-through rate)
| Cost ratio Rank | Style code | Product name | Confirmed Retail price | Cost ratio (%) | Sell-through rate (%) | Supplier code | Supplier name |
|---|---|---|---|---|---|---|---|
| 5 | NDTS5B899IV | Navy cool-touch striped polo | 139,000 | 15.7 | 56.5% | 0000103026 | Hansung A&D Co., Ltd. |
| 7 | NDTS5B404NV | Cool Light navy C/P piqué polo | 159,000 | 16.0 | 47.0% | 0000303572 | Pacific Apparel Ltd |
| 7 | NDTS5B404BR | Cool Light brown C/P piqué polo | 159,000 | 16.0 | 44.2% | 0000303572 | Pacific Apparel Ltd |
| 7 | NDTS5B404BG | Cool Light beige C/P piqué polo | 159,000 | 16.0 | 42.8% | 0000303572 | Pacific Apparel Ltd |
| 7 | NDTS5B404BK | Cool Light black C/P piqué polo | 159,000 | 16.0 | 41.8% | 0000303572 | Pacific Apparel Ltd |
| 7 | NDTS5B404LB | Cool Light light-blue C/P piqué polo | 159,000 | 16.0 | 40.5% | 0000303572 | Pacific Apparel Ltd |
| 7 | NDTS5B404LG | Cool Light light-green C/P piqué polo | 159,000 | 16.0 | 35.8% | 0000303572 | Pacific Apparel Ltd |
| 7 | NDTS5B404BL | Cool Light blue C/P piqué polo | 159,000 | 16.0 | 31.6% | 0000303572 | Pacific Apparel Ltd |
| 1 | NDTS5B897NV | Navy embroidered-logo T-shirt | 109,000 | 12.7 | 26.3% | 0000103026 | Hansung A&D Co., Ltd. |
| 3 | NDTS5B898NV | Navy embroidered-logo polo | 139,000 | 14.9 | 23.3% | 0000103026 | Hansung A&D Co., Ltd. |
💡 ONDO AI insights
- The navy cool-touch polo has a sell-through rate of 56.5%, ranking first. Its 15.7% cost ratio also makes it strong on both margin and turnover.
- Cool Light polos share the same 16.0% cost ratio, but sell-through varies widely by color(navy 47.0% → blue 31.6%). Focus inventory on stronger colors and reduce orders for weaker ones.
- The lowest-cost-ratio (12.7%) embroidered-logo T-shirt has a sell-through rate of just 26.3%. Its margin is strong, but turnover is slow; consider more visibility and promotion.
Recommendation — expand the navy basics line as a core allocation for next season; it performs well on both cost ratio and sell-through.
Evidence: internal product and sales tablesONDO AI verified SQLUnsupported answer blocked ✓
NORDÉ's 2026 SS trench coat is up 23.7% year over year and ranks first in revenue. ← Unsupported guess (fabricated figure)
Type a message...
MCP integration Deep research General search
ONDO AI workspace · demo
How It Works
How internal knowledge becomes an AI answer
Connect terms extracted from documents with data structures in a knowledge graph. Conversational AI refers to this knowledge while people inspect and refine the graph in real time.
Collect
An LLM extracts business terms from internal documents, including image-based files.
Connect
Link data structures and business terms in the graph. User edits remain in place when the source data is reloaded.
Maintain a source of truth
Manage views, columns, joins, and terms in the graph to give AI a consistent reference.
Query
Conversational AI first finds the tables, columns, and joins relevant to a question.
Validate
Compare generated SQL with the graph before execution. Reject rule violations and prompt a rewrite.
Generating and validating SQL are separate steps. Structural errors are checked before execution; business meaning and access rights need their own review.
Scope & Validation
Define the scope with
your own company data
Review database schemas, business terms, and representative questions. Check structural validation, access rights, and operating conditions in the environment where the system will run.
| Inputs | DB schema, business terms, representative questions, access rights, and aggregation rules |
|---|---|
| Structure checks | Compare generated SQL with registered tables, columns, and joins before execution |
| Output and history | Query results, references, rejection reasons, and edit history |
| Deployment | Confirm target databases, supported SQL syntax, schema updates, and execution paths |
| Meaning and access | Review metric definitions, aggregated results, and data access separately |
| Evaluation | Use representative questions, wrong answers, and rejections to assess accuracy, retries, and validation errors |
- Validation depends on the schema and rules registered in the graph. Check that changes to the database are reflected.
- A query may use real columns and still produce a misleading business result. Define correct outcomes and review the answers.
- Review rejections and rewrites for both false acceptances and false rejections.
- Check execution controls and recovery for validator outages, missing permissions, and unsupported syntax.
Structural validation does not guarantee a correct answer or catch every error. Confirm scope and performance with real data, questions, models, and execution conditions.
Related work — Enterprise AI platform
Our experience with internal data queries, access, and usage management informs the knowledge core.
Roadmap
From connected knowledge
to reasoning across relationships
We operate a knowledge graph core. The next step is expanding toward an ontology that can reason over relationships in that knowledge.
- STAGE 0 · Technical validation
Design graph storage and retrieval
Designed a structure for linking schemas and business terms in a graph and reviewing them visually.
- STAGE 1 · Current
Knowledge graph core
We operate schema and business-term linking, SQL validation, live knowledge editing, and improvement suggestions.
- STAGE 2 · Next
Ontology-based reasoning
Explore OWL-based formal reasoning to derive relationships between knowledge. We also plan to assess value ranges, hierarchies, and compatibility with the existing graph.
What would you ask your internal data?
We will use representative questions and data structures to explore how natural-language queries could fit your work.