87 lines
2.6 KiB
Markdown
87 lines
2.6 KiB
Markdown
# Runbook - 06 RAG Vector Classified Docs
|
|
|
|
## Objective
|
|
|
|
Show that a RAG agent retrieves only authorized chunks/documents before sending context to the LLM.
|
|
|
|
## Security Value
|
|
|
|
- Reduces over-retrieval in RAG.
|
|
- Prevents confidential chunks from being sent to the model.
|
|
- Combines vector search with data grants by classification.
|
|
|
|
## Prerequisites
|
|
|
|
- Oracle AI Database with support for `VECTOR`, `TO_VECTOR`, and `VECTOR_DISTANCE`.
|
|
- Database compatible with Oracle Deep Data Security.
|
|
- SQLcl or SQL*Plus.
|
|
|
|
## Before - Vulnerable Environment
|
|
|
|
1. Reset the scenario:
|
|
|
|
```sql
|
|
@scenarios/06-rag-vector-classified-docs/sql/99_reset.sql
|
|
```
|
|
|
|
2. Create chunks and personas without applying data grants:
|
|
|
|
```sql
|
|
@scenarios/06-rag-vector-classified-docs/sql/00_schema.sql
|
|
@scenarios/06-rag-vector-classified-docs/sql/01_seed_data.sql
|
|
@scenarios/06-rag-vector-classified-docs/sql/02_identities.sql
|
|
```
|
|
|
|
3. Simulate the RAG question:
|
|
|
|
```text
|
|
Summarize critical documents about renewals, people, and legal risks.
|
|
```
|
|
|
|
4. Run the vector search:
|
|
|
|
```sql
|
|
@scenarios/06-rag-vector-classified-docs/sql/04_test_queries.sql
|
|
```
|
|
|
|
## Expected Result Before
|
|
|
|
- The search may retrieve `HR_CONFIDENTIAL`, `LEGAL_CONFIDENTIAL`, and `EXECUTIVE_CONFIDENTIAL` chunks.
|
|
- The LLM could receive sensitive context before it even generates an answer.
|
|
|
|
## After - Applying Deep Data Security
|
|
|
|
1. Apply data grants by classification:
|
|
|
|
```sql
|
|
@scenarios/06-rag-vector-classified-docs/sql/03_data_grants.sql
|
|
```
|
|
|
|
2. Run the same search as `nina`, `heitor`, `sofia`, and `carlos`:
|
|
|
|
```sql
|
|
@scenarios/06-rag-vector-classified-docs/sql/04_test_queries.sql
|
|
```
|
|
|
|
## Expected Result After
|
|
|
|
- `nina` retrieves only `PUBLIC` and `INTERNAL` chunks.
|
|
- `heitor` retrieves authorized HR content.
|
|
- `sofia` retrieves authorized legal content.
|
|
- `carlos` retrieves all chunks through the executive role.
|
|
- The RAG layer sends only authorized context to the LLM.
|
|
|
|
## Demo Evidence
|
|
|
|
- Retrieved chunk list before and after protection.
|
|
- Visible classifications by persona.
|
|
- Explanation that enforcement happens before the LLM call.
|
|
|
|
## Official References
|
|
|
|
- Oracle Deep Data Security Guide: https://docs.oracle.com/en/database/oracle/oracle-database/26/ddscg/index.html
|
|
- Create Data Grants: https://docs.oracle.com/en/database/oracle/oracle-database/26/ddscg/create-data-grants.html
|
|
- TO_VECTOR SQL Reference: https://docs.oracle.com/en/database/oracle/oracle-database/26/sqlrf/to_vector.html
|
|
- VECTOR operations in PL/SQL: https://docs.oracle.com/en/database/oracle/oracle-database/26/lnpls/sql-data-types.html
|
|
|