Enterprise RAG
Retrieval That Survives Permissions and Audit
Enterprise RAG, retrieval-augmented generation for organisations, answers questions from your own documents and systems by retrieving the relevant, permitted passages and passing them to a language model, so answers cite their sources and respect access rights. VDF AI ships enterprise RAG as a platform capability: ingestion, embeddings, vector search, permission filtering and generation, all inside your perimeter.
Confluence, SharePoint, Drive, Jira, GitHub, Slack, Notion and SQL, through one governed retrieval layer.
Serving chat today and agent networks tomorrow.
Bring your hardest files. Scans, spreadsheets and contracts welcome.
Where the pilot fell short
- The demo answered questions about a clean folder of PDFs and fell apart on the real document estate
- Someone got an answer sourced from a file they were never allowed to open
- Legal asked where the embeddings live and nobody could say for certain
- Accuracy was a feeling, not a number anyone could defend
What VDF AI does differently
- Retrieval filtered by the source system's permissions at query time: how permission-aware RAG works
- Embeddings, index and generation on infrastructure you control: what private RAG means
- OCR, layout and table handling for scanned documents and spreadsheets
- Accuracy tracked with a labelled evaluation framework
- Delivered through VDF AI Chat and consumed by agents and networks
Retrieval That Works in a Demo
and Retrieval That Works at Work
Demo RAG
- A folder of clean text PDFs uploaded once
- Everyone sees everything
- Embeddings sent to a cloud API per query
- Tables and scans silently dropped
- No idea whether the answer was right
- Deleting a document deletes nothing downstream
Enterprise RAG
- Live sync from wikis, drives, tickets, repositories and databases
- Passages filtered by the reader's entitlements
- Embedding models and vector store on your hardware
- OCR, layout parsing and table extraction in the pipeline
- Recall, faithfulness and citation accuracy measured per release
- Erasure propagated to chunks, vectors and caches with proof
Connect. Retrieve.
Govern.
Connect the systems people already use
OAuth connectors for Microsoft 365, Google Workspace, Atlassian, GitHub, GitBook, Slack, Notion and relational databases keep the index in step with the source and carry its permissions with every document.
Retrieve what is relevant and permitted
Hybrid semantic and keyword search, metadata filters, reranking and optional knowledge-graph retrieval find the right passages. Permission checks run at query time so the answer set is never wider than the reader's access.
Govern every answer
Each response names its sources. Each turn is written to the audit Vault with the retrieved passages, the model and the user. Retention and erasure rules reach the chunks, vectors and caches, and the deletion can be proven.
Three ways to get answers from internal knowledge
| Capability | Cloud RAG API | Enterprise search | VDF AI enterprise RAG |
|---|---|---|---|
| Document content leaves your environment | On every query | Depends on vendor | Never |
| Permission-aware results | Build it yourself | Usually | Yes, from source entitlements |
| Synthesised, cited answers | Yes | Links, not answers | Yes, with sources and audit |
| Scans, tables, multilingual | Varies | Indexing only | OCR, layout and table pipeline |
| Feeds AI agents and networks | Through code | No | Natively |
| Cost at scale | Per query | Per seat | Included in platform capacity |
Eight questions to ask before you buy an enterprise RAG platform
These come from evaluations run by banks, insurers and public bodies. If a vendor cannot demonstrate the answer on your own documents, treat it as a no. The architecture behind each is covered in the private RAG guide and the knowledge graph RAG guide.
- Does retrieval respect document permissions?A user must never receive a passage from a file they cannot open. Permission filtering has to happen at query time, from the source system's entitlements.
- Where do embeddings and the index live?Embeddings are a lossy copy of your documents. If they sit in a third-party cloud, so does the content.
- Can it read the documents you actually have?Scanned PDFs, spreadsheets, tables inside contracts and multilingual policies break naive pipelines. Ask for a test on your worst files.
- How is accuracy measured?Retrieval recall, answer faithfulness and citation correctness on a labelled set from your own corpus, re-run after every change.
- Are answers cited and auditable?Every answer should name its sources, and every turn should be recorded so a disputed answer can be reconstructed.
- What happens when a document is deleted?Erasure must reach the chunks, the embeddings and the caches, and you should be able to prove it.
- Can retrieval feed agents, not only chat?The same governed retrieval should serve agents and networks that act on the answers, not just a chat window.
- What does it cost at your query volume?Per-query cloud RAG APIs scale linearly with usage. Retrieval inside a capacity-licensed platform does not.
RAG by deployment mode, and the products that deliver it
The pipeline is the same everywhere. What changes is where the embedding model runs and where the vector index is stored.
- On-Premises RAGArchitecture, controls and trade-offs for this mode.
- Self-Hosted RAGArchitecture, controls and trade-offs for this mode.
- Sovereign RAGArchitecture, controls and trade-offs for this mode.
- Air-Gapped RAGArchitecture, controls and trade-offs for this mode.
- VDF AI ChatPrivate enterprise chat with permission-aware RAG and per-turn audit.
- VDF AI Data SuiteFine-tune and evaluate models on the same corpus.
- Agentic RAG vs traditional RAGWhen retrieval needs to plan multi-hop queries.
- Semantic search toolsPer-source retrieval tools agents can call.
- AI agent platformThe platform enterprise RAG feeds.
Questions about enterprise RAG
What is enterprise RAG?
Enterprise RAG, short for retrieval-augmented generation, is the practice of answering questions from an organisation's own documents and systems by retrieving the relevant, permitted passages and passing them to a language model, so answers cite sources and respect access rights. It differs from consumer RAG in three ways: permission-aware retrieval, governed data boundaries and per-turn audit.
What is the difference between enterprise RAG and private RAG?
Private RAG describes where the pipeline runs: storage, embeddings, index and generation on infrastructure the enterprise controls, so no document fragments reach a third-party provider. Enterprise RAG describes what the pipeline must do: connect to many systems, filter by permission, handle real documents, cite sources and be measurable. VDF AI delivers enterprise RAG as private RAG by default.
Should we use RAG or fine-tune a model on our documents?
Use retrieval for knowledge that changes, must be cited or is subject to access rights, which covers most enterprise documents. Fine-tune when you need a model to adopt a style, a format or a specialist vocabulary. Many organisations do both: a fine-tuned open-weight model for tone and domain language, with RAG supplying the facts. The Data Suite covers the fine-tuning side.
How do we measure whether enterprise RAG is accurate?
Build a labelled evaluation set of real questions and the passages that answer them, then track retrieval recall, answer faithfulness to the retrieved context and citation correctness. Re-run the set after every chunking, embedding or model change. The evaluation framework guide on this site describes the method and the thresholds regulated teams use.
Why not build enterprise RAG ourselves with a framework and a vector database?
The demo takes a week; the production system takes the year after it. Connectors with OAuth, permission filtering, OCR and table handling, multilingual embeddings, reranking, caching, deletion propagation, evaluation and audit are each a project. A RAG platform ships them and lets your team spend its time on the corpus and the questions.
Evaluate your knowledge stack
Find out how a private RAG and retrieval layer would perform on your data — accuracy, latency, governance, and what to fix before you scale.