# Enterprise RAG knowledge assistant

> RAG (retrieval-augmented generation) is a method where a language model first finds the relevant passages in your company documents and then bases its answer on them. Corrplus builds knowledge assistants that answer from your procedures, contracts, technical documentation and policies, cite the exact source, and never draw on documents the user is not allowed to see.

URL: https://corrplus.net/services/rag-enterprise-knowledge-assistant/
Updated: 2026-10-22
Corrplus Team

## Who it is for

- Companies whose procedures, instructions and policies are scattered across SharePoint, Google Drive or file servers
- Manufacturers and technology firms whose sales, support or field teams spend time hunting for answers in technical documentation
- Legal, finance and HR teams searching through contracts, regulations and internal correspondence
- IT and compliance teams worried that staff upload company documents to personal ChatGPT accounts

## When it is not a fit

- Teams with few documents, or documents nobody keeps up to date. The assistant is only as accurate as its content; tidying the content comes first.
- Numerical analysis and reporting. 'Which dealer sold how much last quarter' is a database query and a report, not RAG.

## What you get

- **Source inventory and access map**: A written record of which document sets are included, who may see what, and which data is masked.
- **Indexing pipeline**: An automated flow that chunks and embeds documents and refreshes the index whenever a document changes (pgvector or Qdrant).
- **Assistant with citations**: Every answer links to the document and section it relies on; when no source is found, it says so plainly.
- **Access control**: A filter tied to your identity system (Microsoft Entra ID, Google Workspace) so users only get answers from documents they are allowed to read.
- **Quality measurement**: A test set of real questions and a regular evaluation report on answer accuracy.
- **User interface**: A web interface or access from Microsoft Teams and Slack, plus an API for your other software when needed.

## How we work

- **Discover**: We collect the questions your teams ask most and where the answers live today. That list sets the scope and becomes the first test set.
- **Design**: We write down the sources, the permission model, the database choice and where data is hosted, reviewed against GDPR and Türkiye's data protection law (KVKK).
- **Build**: We start with one document set, measure and improve against the test set, then add new sources.
- **Run**: We track unanswered questions, flag missing content to document owners and keep the index current.

## Tools we use

pgvector (PostgreSQL), Qdrant, Supabase, OpenAI, Anthropic Claude and Google Gemini models, Azure OpenAI, Open-source embedding models, Microsoft SharePoint and Google Drive connectors, n8n

**Timing:** The first working version on one document set usually ships within 3–6 weeks

**Pricing:** Fixed-scope project quoted after the discovery call; model and hosting costs billed to your own accounts; optional monthly operation.

## What does an enterprise knowledge assistant do?

In most companies the knowledge exists but cannot be found. The safety instruction is in one folder, the product datasheet in another, the latest dealer contract in someone's inbox. A new hire asks a senior colleague, who stops what they are doing to answer.

A knowledge assistant takes that question in plain language, searches the relevant documents and answers with its source: "This comes from the Purchasing Procedure, section 4.2." The user can open the document in one click to check. If the assistant cannot find an answer, it says so instead of guessing.

## How does RAG work?

Documents are split into small chunks, and each chunk is turned into a list of numbers that represents its meaning (an embedding). These are stored in a vector database such as pgvector or Qdrant. When a question arrives, the closest chunks by meaning are retrieved, filtered by the user's permissions and passed to the language model with the instruction to answer only from them.

We use the same method in our own product, Promptor, which indexes a WordPress site's pages and posts and answers visitor questions from that content. Enterprise projects differ in scale and permission model; the principle is the same.

## How does a RAG assistant differ from a chatbot or an LLM integration?

| Aspect | RAG knowledge assistant | [Chatbot](/services/ai-chatbot-development/) | [LLM integration](/services/llm-integration/) |
|---|---|---|---|
| Who uses it | Employees, internally | Website and WhatsApp visitors | Existing users of your software |
| Knowledge source | Large internal document sets, filtered by permission | Public site content, FAQs, product data | The record on screen |
| Critical requirement | Citations and access control | Conversation flow and handover | Cost, speed and consistency |
| Where it lives | Its own interface or inside Teams | Site widget and WhatsApp line | A new button in your existing screens |

A chatbot can use RAG too; the difference is who it serves and which documents it touches. Internal documents never go behind a public chatbot. An LLM integration is not a separate assistant but a feature inside one of your screens. If your support team will use this knowledge base to answer customers, see [AI customer service](/services/ai-customer-service/) as well.

## How do we start?

In the first call we list the questions your teams ask most and where the answers sit. That list is the first test set. We begin with one document set, such as HR policies or product documentation, measure answer quality and then expand. Which documents may enter the assistant is settled at this stage under GDPR and KVKK. For a broader roadmap, start with the two-week [AI discovery sprint](/ai-discovery-sprint/).

## Frequently asked questions

### RAG or fine-tuning a model on our data?

For answering from company knowledge, usually RAG. When a document changes you refresh the index, and every answer shows its source. Fine-tuning suits teaching a model a style or output format; as a way to keep facts current, it is expensive and hard to audit.

### Will the assistant make things up?

The risk is never zero, but it can be reduced. We set the assistant to answer only from retrieved passages and to say when it finds nothing. Each answer carries a source link users can check, and we measure accuracy against the test set regularly.

### Can everyone see every document?

No. Permissions are checked at retrieval, so a passage from a document the user cannot access never reaches the model. Wherever possible, permissions come straight from SharePoint, Google Drive or your identity system.

### Does our data leave the EU or Türkiye?

That depends on the design. The vector database can run on your servers or in a region you choose. For the model, options include regional cloud deployments or open models on your own hardware. Cross-border transfer is assessed with your legal counsel.

### pgvector or Qdrant?

If you already run PostgreSQL or Supabase, pgvector needs no extra system and handles most document volumes. We recommend Qdrant for very large collections, advanced filtering or when the search layer must scale on its own.

### Can it use scanned PDFs and spreadsheets?

Yes, with preprocessing. Scanned documents go through OCR first, and tables are chunked without breaking their structure. Poor scans lower answer quality, so we test with your sample documents during discovery.