Retrieval-augmented generation (RAG) on Oracle Database 23ai is marketed as a single architecture: store vectors next to the source data, run AI Vector Search, ship grounded prompts to an LLM. Oracle's positioning is that AI Vector Search is included in Oracle Database Enterprise Edition at no additional license cost — and that is technically correct. What Oracle's marketing pages do not say is that a working RAG implementation rarely uses AI Vector Search in isolation. It pulls in Partitioning to manage the vector index, Advanced Compression to keep the embeddings footprint manageable, Diagnostics and Tuning Pack to performance-tune the IVF or HNSW index, and Active Data Guard or GoldenGate when the RAG workload needs read replicas. Every one of those is a separately licensed Database Option. Customers running RAG in production without checking the option map are doing the exact thing LMS waits for. This guide walks through the full option footprint of a real RAG deployment on Oracle 23ai, the LMS audit signals to expect, and the architectural moves that keep the licensing footprint clean.
A production RAG implementation on Oracle Database 23ai has four moving parts: a source content repository (often a document store, sometimes a relational schema), an embedding pipeline that converts content into vectors, the vector index inside Oracle Database that handles approximate-nearest-neighbour search, and a prompt-assembly layer that pulls top-k matches into the LLM context. AI Vector Search is the third part. The other three parts have licensing implications too, and they rarely get the attention they deserve in initial architectural reviews.
For most enterprise deployments, the vector index sits next to the operational data. That is the design Oracle markets aggressively — vectors and source data in one place, no ETL, no second store. The economics of that design depend entirely on which Oracle Database Options the schema and index design happen to trigger. See the Oracle Database Licensing Guide for the underlying EE / SE2 / Options framework. The audit pattern below sits inside the broader Oracle audit guide.
Across 14 Oracle 23ai RAG engagements in 2025, the option footprint we mapped was strikingly consistent. Partitioning showed up in 12 out of 14 because vector indexes scale through partitioning. Advanced Compression appeared in 11 because embedding columns are storage-heavy and compression cuts the footprint 60–75%. Diagnostics Pack and Tuning Pack appeared in all 14 because every team eventually runs ADDM, AWR or SQL Tuning Advisor against the RAG workload. Active Data Guard appeared in 7 because RAG read scaling pushes teams toward physical standby with active reads.
The list-price exposure across all six options on a 16-processor cluster is $1.05M plus 22% annual support. After typical Oracle discounting on net-new options, the real-world exposure is $250K to $450K plus support. That is the licensing cost of a RAG architecture that the initial sizing model said was free.
The Oracle Database tracks every option that has ever been touched in DBA_FEATURE_USAGE_STATISTICS. LMS reads this view as part of any audit. Vector Search itself does not create an option-usage flag — Oracle has been careful to keep the base vector functionality off the option-detection trigger. What does create flags is the surrounding usage pattern.
Creating any partitioned object in the same tablespace as the vector schema flags Partitioning. Running an AWR snapshot covering the vector workload flags Diagnostics Pack. Calling DBMS_ADVISOR or DBMS_SQLTUNE against vector queries flags Tuning Pack. Creating an Active Data Guard read replica flags ADG even if no application has connected to it. The LMS audit pattern for AI Vector Search in 2025 is to query DBA_FEATURE_USAGE_STATISTICS, cross-reference USED_NOW_AND_LAST_DETECTED, and follow up on every option whose last detected date sits inside the audit window.
Real LMS pattern: in 2025 audits, LMS asked for DBA_FEATURE_USAGE_STATISTICS snapshots covering the period from initial Oracle 23ai install through the audit notice date. The vector schemas were inspected line-by-line for option usage. Three out of four customers we defended had to pay for at least one option they did not realise was triggered.
The cleanest RAG deployments on Oracle Database 23ai use four principles. First: keep the vector schema separate from any partitioned, compressed or ADG-replicated schema. Second: tune vector queries with EXPLAIN PLAN and manual indexing, not SQL Tuning Advisor — manual tuning leaves no Tuning Pack trail. Third: if you need partitioning, license Partitioning explicitly and document the use case; the cost is small compared to the audit-defence exposure of using it accidentally. Fourth: consider Autonomous Database for RAG workloads under 16 OCPUs — the all-options bundle in ADB removes the option-detection question entirely.
The fourth point matters more than buyers usually appreciate. Autonomous Database licensing includes Partitioning, Advanced Compression, Diagnostics, Tuning, ADG, RAT, Spatial and Graph in the per-OCPU rate. For a 10-OCPU RAG workload, ADB at $1.34 per OCPU-hour list is roughly $117K per year. The equivalent on-prem stack with all the options listed above is closer to $260K per year fully loaded. ADB is the cheaper architecture for RAG below 20 OCPU.
The broader option-detection pattern is covered in the Oracle compliance guide. The compliance review service can map your specific RAG deployment against the option-detection rules and produce a defensible architecture statement before LMS arrives.
Independent, buyer-side analysis. Fixed-fee, 10 business day turnaround. Former Oracle insiders, 25+ years, $1.8B in Oracle spend advised.
AI Vector Search is included in Oracle Database Enterprise Edition and Standard Edition 2 with 23ai at no additional license cost. The vector data type and the vector index types (IVF, HNSW) are part of the base database. Surrounding options - Partitioning, Advanced Compression, Diagnostics Pack, Tuning Pack, Active Data Guard, GoldenGate - are separately licensed if used.
In our engagements, the typical RAG footprint pulls in Partitioning (~ $11,500/proc), Advanced Compression (~ $11,500/proc), Diagnostics Pack and Tuning Pack ($7,500 + $5,000/proc), and sometimes Active Data Guard ($11,500/proc) for read replicas. That is up to $47,000 per processor of option exposure on top of EE.
LMS reads the DBA_FEATURE_USAGE_STATISTICS view, which tracks which optional features have been touched since database creation. AI Vector Search itself does not trigger options, but creating partitioned indexes, using compressed segments, or running ADDM/AWR reports against the vector schema will leave usage trails LMS reviews.
Yes. AI Vector Search is available on SE2 with 23ai. The economics change significantly because SE2 does not support Partitioning, Advanced Compression, Active Data Guard or RAC. For modest-scale RAG, SE2 is a viable cost-saving path.
Yes - Autonomous Database includes all options in the per-OCPU or per-ECPU subscription price. For RAG workloads that fit Autonomous Database's ceiling, ADB is often the simpler commercial choice. The trade-off is data-residency and bring-your-own-customisation flexibility.
Twice a month. Oracle pricing moves, audit-defence tactics, GenAI Service rate changes. Written by former Oracle insiders.
No spam. Unsubscribe any time. Independent — not affiliated with Oracle Corporation.