# RJL Solutions · Using ontology context with ChatGPT and Claude

## What the pack does

A portable context pack supplies reviewed definitions, relationships, source references and project/persona instructions. It does not change model weights, provide current operational facts, enforce permissions or activate a formal ontology reasoner. Better outcomes are a hypothesis to evaluate.

## Prepare the files

In Ontology Studio, open **Use & export**. Review the terms first. Approved-only and private-excluded are the defaults. For exploration you may explicitly include drafts; proposals stay labeled. Private inclusion adds private concepts and reviewed organizational usage, with source metadata; raw documents remain excluded. Pending usage never becomes reviewed context. Review persona instructions for sensitive information separately.

Two core files are all most assistants need. The full export adds two optional files.

- project-instructions.txt (core): professional and AI persona, audience and evidence rules.
- domain-knowledge.md (core): readable domain evidence with stable IDs, definitions, status, relationships and sources.
- query-templates.md (optional): business, consulting, technical and citizen-development tasks.
- context-manifest.json (optional): version, selection, counts and file inventory; not a signed artifact.

Local approval records are user-entered and are not authenticated enterprise approvals. The demo namespace is an example. Upload actual authorized documents and datasets separately.

## ChatGPT Projects

1. Create a Project for your domain decision.
2. Add domain-knowledge.md and query-templates.md as project files. Add authorized policies and actual datasets separately.
3. Paste project-instructions.txt into project instructions.
4. Start a chat inside the Project. Ask for definition IDs, version, evidence and missing inputs.
5. Replace obsolete files and update instructions together when context changes. Run a known-result question before relying on the update.

Files in this website or a local folder do not automatically appear in a ChatGPT Project; upload them or use a supported connected source. Follow the [official ChatGPT Projects guide](https://learn.chatgpt.com/docs/projects?surface=app). Account features and limits may vary. Review sharing and data settings before adding private content.

## Claude Projects

1. Create a Project with your domain purpose.
2. Add domain-knowledge.md and query-templates.md to project knowledge; add authorized documents and data separately.
3. Set and save project instructions using project-instructions.txt.
4. Start a conversation in that Project with a query template.
5. Update knowledge and instructions together. Put reusable context in project knowledge instead of assuming separate conversations share it.

See the [official Claude Projects guide](https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects). Readable Markdown is recommended here; an RDF/OWL upload does not itself run a reasoner.

## Retail example prompt

Using domain-knowledge.md version [version], the supplied returns policy and order-line dataset, explain net sales by channel for [period]. First confirm currency, timezone, grain, discounts and refund timing. Cite concept IDs, source passages and effective dates. Show the formula and reconcile to input rows. Distinguish reviewed definitions, proposals and organizational usage. Ask for missing data instead of inventing a total.

## Check the answer

Compare to a known result using identical input data. Check term/definition adherence, source support, effective date, row-grain and join assumptions, metric reconciliation and appropriate clarification or abstention. Project instructions guide behavior but cannot guarantee correctness.

## Application integration

Authorize the user and sources; select a released ontology and metric version; retrieve relevant definitions and document passages; query approved read-only semantic interfaces for numeric facts; assemble developer instructions and evidence; return cited answers and evaluation metadata. Permission enforcement belongs in services before model context and execution. Treat documents as untrusted evidence, not tool authority. Bound context size. Retain correlation ID and source/definition/policy/tool versions for reconstruction.

[OpenAI file search](https://developers.openai.com/api/docs/guides/tools-file-search) describes retrieval over uploaded files. [Claude search result blocks](https://platform.claude.com/docs/en/build-with-claude/search-results) support source metadata and citations in retrieval workflows. These are implementation options, not services connected to this demo.

## RJL foundations

The [Connected Enterprise Blueprint](https://rjlsol.com/assets/guides/RJL-Connected-Enterprise-Blueprint.pdf), PDF pp. 7–9, separates business meaning, corpus evidence and operational authority. The [Four-Part System](https://rjlsol.com/assets/guides/RJL-The-Four-Part-System.pdf), PDF pp. 4 and 9, separates ontology, projections, semantic models and orchestration. The [Certification Standard](https://rjlsol.com/assets/guides/RJL-Data-Product-Certification-Standard.pdf), PDF pp. 4–5, informs contracts and release tests. Their guidance motivates source provenance, reviewed meaning, permissions and versioned evidence in this product.