Oracle AI Agent Knowledge Sources
An Oracle AI agent knowledge source is a set of documents you upload to an agent so it can answer from your trusted content through semantic search rather than from the model's general training. In Oracle AI Agent Studio you attach a knowledge source, the documents are indexed for meaning, and at run time the agent retrieves the most relevant passages and grounds its response in them — keeping answers tied to your policies and reference material rather than inventing one.
This page covers how to configure knowledge sources, keep their content current, control who can see grounded content, and validate that responses are actually grounded in the source you provided. Uploading documents for semantic search is generally available with AI Agent Studio (release 25C). For the wider Studio overview, see the Oracle AI Agent Studio page.
Last reviewed: 19 July 2026
Scope note — knowledge sources vs document tools. A knowledge source is reference content the agent reads from: documents uploaded for semantic search so the agent can ground its answers in them. A document tool is an action the agent takes on documents — an operation it invokes as part of a task. They are easy to confuse because both involve documents, but one is passive grounding and the other is an active capability. This page owns knowledge-source configuration, freshness, access control, and grounding validation. For the tools an agent uses to act on documents, see Oracle AI Agent Document Tools. The broad product overview stays on AI Agent Studio.
What a Knowledge Source Is
A knowledge source is trusted content you make available to an agent so its answers are drawn from your material instead of the language model's general knowledge. In Oracle AI Agent Studio you add documents to an agent; those documents are indexed so they can be searched by meaning, not just keyword. When a user asks a question, the agent retrieves the passages most relevant to it and uses them as the basis for its response — the pattern industry practice calls retrieval-augmented generation, and the officially supported way to point a Fusion agent at your own content.
The value is specificity and trust. A general model can describe how expense policies usually work; a grounded agent can answer how your expense policy works, citing the paragraph that says so. The knowledge source also sets the ceiling on answer quality: an agent can only be as accurate, current, and complete as the documents behind it.
Knowledge sources work alongside the agent's instructions and prompts, which tell the agent when to lean on its grounded content. Instructions set the behaviour; the knowledge source supplies the facts — and both must be right for the agent to be trustworthy.
Source Types & Configuration
Oracle's officially documented knowledge-source pattern is uploading documents to an agent for semantic search. The configuration decisions below shape how well that grounding works. Availability reflects the verified Oracle facts as of the review date; confirm current behaviour against Oracle documentation before you build.
| Configuration dimension | What you decide | Why it matters | Availability |
|---|---|---|---|
| Document upload | Which documents to attach as the agent's knowledge | Defines the entire universe the agent can answer from | GA (25C) |
| Content scope | Whether one broad set or several focused sets per topic | Narrow, on-topic content retrieves more accurately | GA (25C) |
| Document quality | Clean, structured, de-duplicated source files | Conflicting or stale copies degrade every answer | GA (25C) |
| Agent association | Which agent(s) the knowledge source is attached to | Content should reach only agents that should use it | GA (25C) |
| Instruction alignment | How prompts tell the agent to use the grounded content | Grounding only helps if the agent is told to prefer it | GA (25C) |
| Refresh procedure | How and when documents are re-uploaded when they change | Stale content produces confidently wrong answers | GA (25C) |
| Access boundary | Who can reach an agent grounded in sensitive content | Grounding must not expose content a user shouldn't see | GA (25C) |
Do not assume file-format limits, size caps, or index refresh timings beyond what Oracle's documentation states — confirm current specifics with Oracle before committing to a design.
Freshness & Currency
A knowledge source is a snapshot of the documents at the moment they were indexed. Unless it is refreshed, the agent keeps answering from that snapshot even after the underlying policy or reference has changed. Because a grounded agent answers confidently, a stale source is more dangerous than no source — it produces answers that sound authoritative but are out of date.
Freshness is an ownership problem, not a technical one. Every knowledge source should have a named owner, a defined review cadence, and a re-upload procedure triggered whenever the material is revised. When a policy is superseded, the old document must be removed as well as the new one added — leaving both in the source means the agent can retrieve and cite the wrong one.
Currency also intersects with Oracle's quarterly cadence. When a business process changes because of a Fusion update, the reference documents that describe it need to change too, and the source needs re-indexing. Treat a refresh as a standard step in your change process, the same way you would update a runbook.
Access Control — An Agent Must Not Surface What a User Shouldn't See
The single most important governance rule for knowledge sources is this: an agent must not surface, through a grounded answer, content that the person asking is not entitled to see. Grounding creates a new path to information. If sensitive material is uploaded to an agent that a broad population can reach, the agent can effectively read that material aloud to anyone who asks the right question — even if those users could never open the source document directly.
Two controls prevent this, applied together. First, be deliberate about what goes into a knowledge source — restricted or personal data should not be uploaded to a broadly available agent at all. Second, restrict who can reach an agent grounded in sensitive content, so access to the agent matches the intended audience: an agent grounded in confidential finance material should be reachable only by the finance population entitled to it.
This is also where document tools and knowledge sources must be reasoned about separately: a tool that acts on documents is governed by what that action can reach, while a knowledge source is governed by what content it contains and who can query the agent.
Oracle positions Fusion agentic capabilities as operating within existing role-based permissions, policies, and approval frameworks. Treat that as the baseline, not the finish line: you still decide, per knowledge source, whether the content is appropriate for the agent's audience.
Grounding & Why It Reduces Hallucination
Hallucination is what happens when a language model produces a fluent, plausible answer that is not supported by any real source. A model with no grounding fills gaps with statistically likely text — fine for casual questions, unacceptable for policy, compliance, or financial answers. Grounding narrows the model's job: instead of recalling an answer, it summarises passages retrieved from your knowledge source.
This reduces hallucination in two ways: it gives the model correct, specific material to work from, and it gives you something to check the answer against. A grounded answer can be traced back to the passage it came from; an ungrounded one cannot. That traceability is why the knowledge source is the highest-leverage thing you configure.
Grounding reduces hallucination — it does not eliminate it. An agent can still retrieve the wrong passage or answer confidently when the source does not cover the question. That residual risk is exactly why grounded responses must be validated rather than trusted on sight.
Validating Grounded Responses
Configuring a knowledge source is the start; proving the agent actually answers from it is the discipline that makes it safe to deploy. Validation of grounded responses asks a few precise questions: did the agent answer from the knowledge source rather than from general knowledge, did it retrieve the correct passage, and did it correctly decline or defer when the source did not contain the answer?
What grounded-response validation checks
- Faithfulness: the answer is supported by the retrieved passage, with nothing added that the source does not say.
- Retrieval accuracy: the passage the agent used is the right one for the question asked.
- Coverage: questions the source should answer are answered, not missed.
- Graceful gaps: when the source does not cover a question, the agent declines or escalates rather than inventing an answer.
- Freshness in practice: the agent reflects the current document, not a superseded version still in the index.
- Access boundaries: a user cannot elicit content through the agent that they are not entitled to see.
Grounded-response validation is one slice of the wider evaluation discipline for agents. The methodology — building test sets, scoring answers, and re-running them across changes — belongs on the Oracle AI agent evaluation page; this page focuses specifically on the knowledge-source dimension of it. Oracle's own words for these activities are monitoring, validation, testing, and debugging — there is no separately branded Oracle "evaluation" product, so treat structured evaluation as a practice you apply, not a module you enable.
Knowledge-Source Configuration Checklist
A practical checklist for standing up and maintaining a knowledge source you can trust. Work through it per source, not per agent, since one agent may draw on several.
| # | Checklist item | What "done" looks like |
|---|---|---|
| 01 | Define the purpose | The questions this source should answer are written down and bounded |
| 02 | Curate the documents | Authoritative, current, de-duplicated files — no draft or superseded copies |
| 03 | Classify the content | Sensitivity of each document is known before it is uploaded |
| 04 | Scope the source | Focused topic sets rather than one undifferentiated pile |
| 05 | Attach to the right agent | Only agents that should use the content are associated with it |
| 06 | Align the instructions | Prompts tell the agent to prefer grounded content and defer when it lacks it |
| 07 | Set the access boundary | Who can reach the agent matches who may see the content |
| 08 | Assign an owner | A named person is accountable for this source's accuracy |
| 09 | Define the refresh cadence | A review schedule and re-upload trigger are documented |
| 10 | Build a validation set | Representative questions with known-good answers exist to test against |
| 11 | Test grounding & gaps | Answers are faithful; out-of-scope questions are declined, not invented |
| 12 | Re-validate on change | The validation set re-runs after any document or quarterly-update change |
Security & Configuration Notes
A few principles keep knowledge-source risk manageable. Minimise what you upload: a source should contain the content needed to answer its questions and nothing more, because everything in it is potentially retrievable. Classify before you upload, so a sensitivity decision is never made by accident. And keep access to the agent aligned with the content, so grounding never becomes an unintended disclosure path.
On the configuration side, drift is the quiet failure mode. Knowledge sources, the instructions that reference them, and the access model around them can differ between test and production, and a difference in any one changes the answers an agent gives — a grounded response that is correct in test can be wrong in production. Comparing that configuration across environments is where SyntraFlow Configuration Intelligence can complement the picture: it is built to surface configuration drift, and knowledge-source and agent configuration is a natural extension of that discipline.
Finally, do not assume behaviour Oracle has not documented. File-format handling, indexing timings, retention, and size limits should be confirmed against Oracle's current documentation rather than inferred.
How SyntraFlow Helps
SyntraFlow's own AI-driven testing engine is a separate capability from Oracle's AI — it is used to help organisations assess and validate Oracle agents, including how they use knowledge sources. The framing below is deliberately hedged; specifics are confirmed at assessment.
Grounding assessment
SyntraFlow helps organisations assess whether an agent answers from its knowledge source rather than from general model knowledge.
Validation-set design
SyntraFlow can be configured to help build representative question sets with known-good answers for a given knowledge source.
Freshness regression
The SyntraFlow roadmap can support re-running grounded-answer checks after documents change or a quarterly update lands.
Access-boundary checks
SyntraFlow can be configured to probe whether an agent surfaces content beyond a test user's entitlement.
Config-drift visibility
SyntraFlow can connect configuration comparison with agent testing so drift in sources or access is surfaced, not assumed.
Evidence for sign-off
SyntraFlow helps organisations capture grounded-response results as evidence for governance and audit review.
A note on capability. SyntraFlow's AI-testing engine is distinct from Oracle's AI Agent Studio. The items above describe how SyntraFlow can be configured or extended to help validate knowledge-source grounding; they are not claims that every Oracle AI feature is supported today. Scope is confirmed during an assessment rather than assumed here.
Official Oracle References
The Oracle sources relevant to knowledge sources and AI Agent Studio, verified as of the review date. These are external Oracle links; confirm current details against them before you build.
- Oracle Fusion AI — product page: oracle.com/applications/fusion-ai
- AI Agent Studio — key capabilities (including knowledge sources): docs.oracle.com — key-capabilities
- 25C GA readiness (common): docs.oracle.com — 25C readiness
Last reviewed: 19 July 2026. Availability of uploading documents for semantic search: generally available with AI Agent Studio, release 25C.
Related Oracle AI Pages
Knowledge sources sit within the broader agent-building picture. Go deeper on adjacent topics:
Oracle AI Agent Studio ⭐
The Studio overview and building blocks.
AI Agent Document Tools →
Actions an agent takes on documents.
Prompts & Instructions →
How the agent is told to use grounding.
AI Agent Evaluation →
The wider grounded-response evaluation method.
Oracle AI Hub →
The full Oracle AI content hub.
Configuration Intelligence →
Compare agent and source config across environments.
Frequently Asked Questions
What is an Oracle AI agent knowledge source?
▼
It is trusted content you upload to an agent in Oracle AI Agent Studio so the agent can answer from your documents through semantic search. At run time the agent retrieves the most relevant passages and grounds its response in them, keeping answers tied to your material rather than the model's general knowledge. Uploading documents for semantic search is generally available with AI Agent Studio, release 25C.
How is a knowledge source different from a document tool?
▼
A knowledge source is content the agent reads from to ground its answers; a document tool is an action the agent performs on documents as part of a task. One is passive grounding, the other is an active capability. This page owns knowledge-source configuration and grounding; the tools that act on documents are covered on Oracle AI Agent Document Tools.
Does grounding stop an agent from hallucinating?
▼
Grounding reduces hallucination but does not eliminate it. Giving the model correct, specific passages to work from — and something to check the answer against — makes answers far more reliable, but the agent can still retrieve the wrong passage or answer when the source does not cover the question. That residual risk is why grounded responses should be validated rather than trusted on sight.
How do I keep a knowledge source current?
▼
Give each source a named owner, a review cadence, and a re-upload procedure triggered whenever the source material changes. Remove superseded documents as well as adding new ones, so the agent cannot retrieve an old version. Treat a refresh as a standard step in your change process, including after Oracle quarterly updates that alter the underlying business process.
Can an agent expose content a user shouldn't see?
▼
It can if you let it. Grounding creates a new path to information, so sensitive content uploaded to a broadly reachable agent can effectively be read aloud to anyone who asks the right question. Prevent this by being deliberate about what goes into a source and by restricting who can reach an agent grounded in sensitive content, so access to the agent matches the intended audience for the content.
How do I validate that responses are actually grounded?
▼
Build a representative set of questions with known-good answers and check faithfulness, retrieval accuracy, coverage, and graceful handling of gaps — the agent should decline or escalate when the source does not cover a question. Re-run the set after documents or quarterly updates change. The broader methodology lives on the AI agent evaluation page.
Is there an Oracle "evaluation" product for grounding?
▼
Oracle's documented words for these activities are monitoring, validation, testing, and debugging — there is no separately branded Oracle "Evaluation" module. Treat structured evaluation of grounded responses as a practice you apply rather than a product you switch on, and confirm any specific Oracle capability against current documentation.
How can SyntraFlow help with knowledge-source grounding?
▼
SyntraFlow's AI-testing engine is separate from Oracle's AI. SyntraFlow helps organisations assess whether an agent answers from its knowledge source, can be configured to build validation sets and probe access boundaries, and can connect configuration comparison with agent testing so drift is surfaced. These are hedged capabilities scoped at assessment, not a claim that every Oracle AI feature is supported today.
Ground Your Oracle AI Agents in Trusted Content
Assess how your Oracle AI agents use knowledge sources, where grounding could drift, and how to validate responses before they reach users. See how SyntraFlow can support an Oracle AI agent knowledge-source review.