> ## Documentation Index > Fetch the complete documentation index at: https://docs-eos.fpt.ai/llms.txt > Use this file to discover all available pages before exploring further. # What is Enterprise Data Platform > A shared data layer where company documents are ingested, processed and served to every AI agent - Governed, cited and measurable. Welcome to the **FPT.AI Enterprise Data Platform (EDP)** documentation. An agent only answers correctly when it reads the right documents. EDP is where a company's documents - Processes, policies, legal texts, contracts, templates - Are brought into one place, turned into searchable knowledge, and served to every agent system through one common interface. It is not a document Q\&A chatbot; it is the data layer that sits behind all of your agents. ## Why a separate data layer Most companies feed documents to AI like this: Each team builds its own retrieval for its own agent, documents are copied into many places, and nobody knows which version is current. When an agent answers wrongly, nobody knows where it went wrong. EDP changes that: | Companies today | Companies on EDP | | - | - | | Each agent has its own copy of the documents | One source, many knowledge bases, shared by every authorized agent | | A wrong answer could be the document, the chunking or the model - Nobody can tell | See exactly which passages the agent received before it answered | | Changing how documents are processed means fixing every agent | Change it once in the knowledge base; every connected agent benefits | | No one knows which agent is using which knowledge base | Each KB has its own endpoint and keys; every call is recorded in the KB's call log | ## How it differs from building RAG per agent | | RAG built per agent | Enterprise Data Platform | | - | - | - | | Documents | Uploaded separately into each agent | Ingested once into a Data source, folder tree preserved | | Processing | Each team writes its own chunking and OCR | Ready-made pipelines per file type, tunable in the UI | | Connection | Tied to one framework | MCP and REST API - Claude, Cursor, VS Code, Dify, LangGraph or custom agents | | Quality | Checked by gut feeling | Retrieval test with diagnosis, golden question sets, live quality tracking | | Control | Anyone with the code can read everything | Per-KB API keys or OAuth, revocable, with call log | The quickest way to tell them apart: Agents do the work. EDP is the company library - Where documents are organized, kept current, and lent to the right people. ## Five building blocks Data ingestion and Knowledge are the foundation - Without clean documents, every answer downstream is wrong. Quality, Connection and Administration are what bring that knowledge into real operations. | # | Block | Answers the question | | - | - | - | | 1 | Data ingestion | Where documents come from, and how they are organized and tagged with metadata | | 2 | Knowledge | Where agents search, and what they are allowed to search | | 3 | Quality | Does the agent get the right passages, and did the last change help or hurt | | 4 | Connection | How agents connect | | 5 | Administration | Who can do what, and what information is attached to documents | ### 1. Data ingestion Upload files or whole folders into a **Data source** - PDF, DOCX, XLSX, PPTX, TXT, MD, HTML. The folder tree is preserved and becomes filterable metadata. The pipeline picks a method per file type: DOCX split by headings, scanned PDFs get OCR automatically, spreadsheets split by rows with column headers kept. ### 2. Knowledge A **Knowledge base** groups one or more sources into something agents can search. Each KB has its own embedding model, retrieval method (Vector, BM25 or Hybrid + rerank) and endpoints. For example: A "Legal documents" KB turns on *only return documents in effect*; an "Accounting procedures" KB turns on *expand to parent chunk* so the agent receives a whole Article instead of half a sentence. ### 3. Quality **Retrieval test** shows exactly which passages an agent will receive - The first thing to check when an agent answers wrongly. **Quality evaluation** keeps a golden question set for checking a KB, and shows the questions agents ask that the KB cannot answer yet. ### 4. Connection Every KB has its own MCP URL and REST URL. Agents connect with OAuth (the user signs in and the agent reads with that user's own permissions) or with an API key issued for that KB. ### 5. Administration A metadata schema shared across the workspace, member roles, and ownership transfer when someone leaves. ## Key concepts | Concept | Meaning | | - | - | | Data source | Where documents enter the platform, by uploading files and folders | | Knowledge base | What agents retrieve from, and the unit access is granted on. Always built from one or more sources | | Pipeline | The steps that parse, chunk, enrich, embed and index documents | | Chunk | A passage after splitting - What the agent actually receives | | Metadata | Information attached to a file (confidentiality, effective date, department…) used to filter at retrieval time | | Endpoint | The KB's MCP / REST URL, used with an API key or OAuth sign-in | > [!WARNING] > > **There are no document-level permissions yet.** Every workspace member can read uploaded files, and anyone with a KB's API key can query everything in it. Do not upload confidential documents at this stage. ## Who EDP is for | Audience | Main value | | - | - | | Agent builders | No need to write a retrieval layer; connect a KB over MCP in minutes | | Business document owners | Upload, organize folders, add metadata - No need to know about vectors or chunks | | Data & AI teams | Tune retrieval methods, review chunks and track quality | | IT & Security | Per-KB keys, OAuth, call log, centralized metadata schema | ## Where to start **1. Create a data source** Go to **Data sources** and upload a folder of business documents. Keep the folder tree so you can filter on it later. **2. Create a knowledge base** Select the new source, keep the **Automatic per file** pipeline, and write a description telling agents when to call this KB. **3. Test retrieval** Ask the questions real users will ask and check the returned passages before handing the KB to an agent. **4. Connect an agent** Copy the KB's MCP URL into your agent, sign in with OAuth or issue an API key - The agent starts searching right away. See each step in detail on the [Quick start](https://docs-eos.fpt.ai/en/user-guide.md) page, and a detailed description of each screen in the **Screens & features** group. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.