RAG Topics — Overview

A Topic is a private AI chatbot that answers questions from your documents. It is the central concept of AI Knowledge Fabric. You upload content, and the platform makes it searchable. A language model then answers questions from that content, instead of relying only on what it learned in training.

Each row is one Topic. The status badge shows whether the Topic is running. The icons on the right start the Topic, stop it, and open its settings. A Topic’s name becomes a link to its chat once it is Running.

1. What problem it solves

A general-purpose language model knows nothing about your internal HR policies, your product manuals, or last quarter’s reports. If you ask about them, it either refuses or invents an answer.

RAG — Retrieval-Augmented Generation — fixes this. It searches your documents first. Then it hands the matching passages to the model together with the question:

flowchart TD Q["Question:
'vacation days after 5 years?'"] --> E["1 · Turn the question
into numbers (embedding)"] E --> V["2 · Search
finds passages"] V --> L["3 · Question + passages
→ language model"] L --> A["Answer
+ source: hr-policy.pdf"]

What this means in practice:

	Plain chatbot	RAG Topic
Knows your documents	No	Yes
Shows sources	No	Yes — answers from your documents list the files they used
Answer when nothing is found	Invents something	Shows the configured no-answer response
Adding knowledge	Retraining	Uploading a file
Data location	The provider’s	Your infrastructure (unless you deliberately choose a cloud model)

Plain chatbot

RAG Topic

Knows your documents

Yes

Shows sources

Yes — answers from your documents list the files they used

Answer when nothing is found

Invents something

Shows the configured no-answer response

Adding knowledge

Retraining

Uploading a file

Data location

The provider’s

Your infrastructure (unless you deliberately choose a cloud model)

2. Why you would want several Topics

Each Topic is separate: its own documents, its own models, its own prompts, its own access permissions. That separation is the reason to create more than one:

Separate subject areas. An HR Topic cannot accidentally answer from technical manuals. It usually gives better answers than one Topic that holds everything, because the search has less noise to sort through.
Separate audiences. Access is granted per Topic and per role. So a Topic with salary data can be limited to a single role.
Different models per purpose. A public FAQ can use a small local model. A demanding analysis Topic can use a large cloud model.
Different tone. Each Topic has its own prompts. One can be short and factual, another patient and explanatory.

3. Architecture — what actually runs

This is where AI Knowledge Fabric differs from most chat products: the management application is not the chatbot. It starts and manages containers. When you press Start on a Topic, the platform starts one container for it — the Topic service.

flowchart TD APP["Management application
(this web UI)"] APP -- "starts, per Topic" --> TOPIC subgraph TOPIC ["Per Topic"] direction TB RAG["Topic service
(one container)"] end TOPIC -- "reaches out to" --> DEPS subgraph DEPS ["Registered by an administrator"] direction LR AI["AI servers
(models)"] DOCLING["Docling servers"] VDB["Vector databases"] MCP["Other Topics and
MCP tool servers"] end

Models are not part of the Topic. The embedding model and the language model run on the AI servers registered under Admin → AI Servers. The Topic calls them.
The vector database is not part of the Topic either. Every Topic needs one registered under Admin → Vector Databases. Your documents are stored there, so they survive a stop and start of the Topic.

Three things follow from this design and explain most surprises:

Most settings reach the Topic when it starts. Models, prompts, timeouts and search settings are handed to the container at start — not with each question. Until the next restart, a running Topic keeps the old settings. On the Configuration tab, the app therefore offers Save & Restart for such a change. A few settings apply without a restart, for example the Chunking Strategy and the Docling Settings, which are used with each upload, and the connections on the MCP Interfaces tab.
Each Topic has its own port (RAG Service Port, set in the first step of the wizard). With Docker, the Topic’s own interface is reachable on this port. This is what makes the REST interface possible.
Docker (or Kubernetes) must be available. Without it, a Topic goes from Starting… to Error.

A valid license is needed to create and to start Topics.

Lifecycle states

Status Meaning

Status	Meaning
`Created` / `Configured`	Saved in the database, nothing running yet
`Starting…`	The container image is downloaded and the container is started. The Topic is `Running` only when it answers its own health check
`Running`	Ready for uploads and chat
`Stopped`	Container stopped. Configuration and stored data are kept
`Error`	The start failed — most often Docker is unreachable or a server is misconfigured. A running Topic that stops answering also shows `Error`. It returns to `Running` by itself when it answers again

Created / Configured

Saved in the database, nothing running yet

Starting…

The container image is downloaded and the container is started. The Topic is Running only when it answers its own health check

Running

Ready for uploads and chat

Stopped

Container stopped. Configuration and stored data are kept

Error

The start failed — most often Docker is unreachable or a server is misconfigured. A running Topic that stops answering also shows Error. It returns to Running by itself when it answers again

4. The building blocks you configure

Block	Role	Configured under
Embedding Model	Turns documents and questions into numbers that the search compares. Decides how well the search finds matching passages	AI Servers; chosen in step 2 of the wizard
Large Language Model	Writes the answer from the question and the passages found	AI Servers; chosen in step 2 of the wizard
Vector database	Stores the searchable form of your documents. Required: a registered database	Vector Databases; chosen in step 3 of the wizard
Docling server	Turns PDF, DOCX and PPTX files into text before they are stored	Docling Servers
Image Describer Model	Describes pictures during upload, so their content can be found. Can be its own model, or the Topic’s Language Model. Either way, the model must be able to read images	AI Servers
Retrieval Strategy	How passages are found and sorted	Per Topic, step 4 of the wizard
Prompts	The instructions sent to the models: the shared system prompt, the Topic’s Custom System Prompt, and every other prompt of the Topic	Admin → LLM Configuration, and per Topic (Configuration and Prompts tabs)

Block

Role

Configured under

Embedding Model

Turns documents and questions into numbers that the search compares. Decides how well the search finds matching passages

AI Servers; chosen in step 2 of the wizard

Large Language Model

Writes the answer from the question and the passages found

AI Servers; chosen in step 2 of the wizard

Vector database

Stores the searchable form of your documents. Required: a registered database

Vector Databases; chosen in step 3 of the wizard

Docling server

Turns PDF, DOCX and PPTX files into text before they are stored

Docling Servers

Image Describer Model

Describes pictures during upload, so their content can be found. Can be its own model, or the Topic’s Language Model. Either way, the model must be able to read images

AI Servers

Retrieval Strategy

How passages are found and sorted

Per Topic, step 4 of the wizard

Prompts

The instructions sent to the models: the shared system prompt, the Topic’s Custom System Prompt, and every other prompt of the Topic

Admin → LLM Configuration, and per Topic (Configuration and Prompts tabs)

The embedding model is fixed once the Topic is created. The stored data only makes sense to the model that created it. The same is true for the vector database, the Chunk Size and the Chunk Overlap. To change one of them, create a new Topic and upload the documents again — so decide before the first upload.

5. What a Topic can do

Feature	What you notice	Where you set it
Streaming answers	The answer appears word by word while it is written. You can stop waiting for an answer	Always on
Routing	Messages that need no search — thanks, small talk, "shorter please" — are answered directly, which is faster. A message is answered without a search only if the model is sure enough (Router Confidence Threshold, default 0.9). On by default	Route before retrieval, wizard step 4 and Configuration tab
Keyword Enrichment	Only with Retrieval Strategy HYBRID. ALGORITHMIC (default) takes keywords from the text: fast. LLM lets the Language Model write keywords, also other words with the same meaning: better matches, but slower uploads	Wizard step 4 and Configuration tab
Reranking	LLM lets the Language Model sort the found passages, so the best ones are used. More exact answers, but each question takes longer. Off (NONE) by default	Wizard step 4 and Configuration tab
Chunking Strategy	RECURSIVE (default) cuts documents at paragraphs, sentences and words. SEMANTIC cuts the same way, then lets the Language Model add a short note to each piece about where it belongs. Pieces are found better, but uploads take much longer (one model call per piece)	Wizard step 5 and Configuration tab; applies to the next upload
Structured data import	JSON, XML, YAML, TOML, CSV and TSV files are split into records. Flatten (default) stores each record as "key: value" lines: fast. Narrate lets a model rewrite each record as a short text: questions in normal language find it better, but large files take much longer	Wizard step 2 and Configuration tab
Pictures	Pictures in documents can be described by a vision model, so their content can be found. They can also be stored and shown inside answers	Image Handling, wizard step 2 and Configuration tab
Docling Settings	Advanced settings for how Docling reads PDF and Office files, for example text recognition in scans. Used with every upload; a change applies to the next upload	Wizard step 2 and Configuration tab, next to the Docling server
Chat attachments	Users can attach a file to one question. It is used for that answer only and is not added to the Topic’s documents. Text files are allowed by default	Wizard step 2 and Configuration tab
Other Topics as tools	The Topic can ask other Topics (also GraphRAG topics) while it answers. Their sources and pictures are shown in the answer too. They open only if you have access to that other Topic	MCP Interfaces tab
Prompts	Every instruction the Topic sends to a model can be changed for this Topic	Prompts tab, see Prompts
Public address	Where the Topic is reachable from outside. Links that the Topic hands out to other programs use it	Admin → Public Endpoints, or Public Address Override on the Configuration tab
Reset Topic Data	Deletes the list of uploaded files and all chats of the Topic, for example after testing. The documents stay in the Topic and can still be found	Actions tab
DEV reports	For developers and support: detailed reports of single uploads and answers, with every step and model call. Needs a special permission	DEV Reports tab

Feature

What you notice

Where you set it

Streaming answers

The answer appears word by word while it is written. You can stop waiting for an answer

Always on

Routing

Messages that need no search — thanks, small talk, "shorter please" — are answered directly, which is faster. A message is answered without a search only if the model is sure enough (Router Confidence Threshold, default 0.9). On by default

Route before retrieval, wizard step 4 and Configuration tab

Keyword Enrichment

Only with Retrieval Strategy HYBRID. ALGORITHMIC (default) takes keywords from the text: fast. LLM lets the Language Model write keywords, also other words with the same meaning: better matches, but slower uploads

Wizard step 4 and Configuration tab

Reranking

LLM lets the Language Model sort the found passages, so the best ones are used. More exact answers, but each question takes longer. Off (NONE) by default

Wizard step 4 and Configuration tab

Chunking Strategy

RECURSIVE (default) cuts documents at paragraphs, sentences and words. SEMANTIC cuts the same way, then lets the Language Model add a short note to each piece about where it belongs. Pieces are found better, but uploads take much longer (one model call per piece)

Wizard step 5 and Configuration tab; applies to the next upload

Structured data import

JSON, XML, YAML, TOML, CSV and TSV files are split into records. Flatten (default) stores each record as "key: value" lines: fast. Narrate lets a model rewrite each record as a short text: questions in normal language find it better, but large files take much longer

Wizard step 2 and Configuration tab

Pictures

Pictures in documents can be described by a vision model, so their content can be found. They can also be stored and shown inside answers

Image Handling, wizard step 2 and Configuration tab

Docling Settings

Advanced settings for how Docling reads PDF and Office files, for example text recognition in scans. Used with every upload; a change applies to the next upload

Wizard step 2 and Configuration tab, next to the Docling server

Chat attachments

Users can attach a file to one question. It is used for that answer only and is not added to the Topic’s documents. Text files are allowed by default

Wizard step 2 and Configuration tab

Other Topics as tools

The Topic can ask other Topics (also GraphRAG topics) while it answers. Their sources and pictures are shown in the answer too. They open only if you have access to that other Topic

MCP Interfaces tab

Prompts

Every instruction the Topic sends to a model can be changed for this Topic

Prompts tab, see Prompts

Public address

Where the Topic is reachable from outside. Links that the Topic hands out to other programs use it

Admin → Public Endpoints, or Public Address Override on the Configuration tab

Reset Topic Data

Deletes the list of uploaded files and all chats of the Topic, for example after testing. The documents stay in the Topic and can still be found

Actions tab

DEV reports

For developers and support: detailed reports of single uploads and answers, with every step and model call. Needs a special permission

DEV Reports tab

The Topic’s tabs

The settings icon in the Topic list opens the Topic’s page with eight tabs: Configuration · LLM Overrides · Prompts · Data Upload · MCP Interfaces · Actions · Logs · DEV Reports. See Configuration and functions.

6. MCP — giving a Topic tools

MCP (Model Context Protocol) is a standard for connecting AI services to tools. A running Topic can be connected to MCP servers. Its language model can then call those tools while it answers, instead of relying only on its own documents.

Two uses:

Use	What it achieves
Topic → Topic	Every running Topic and GraphRAG topic can be used as a tool by another Topic. So you can keep knowledge bases separate and still answer questions that need more than one of them
External tool servers	Web search, ticket systems, APIs — anything with an MCP server. You enter its address yourself in MCP Server URL

Use

What it achieves

Topic → Topic

Every running Topic and GraphRAG topic can be used as a tool by another Topic. So you can keep knowledge bases separate and still answer questions that need more than one of them

External tool servers

Web search, ticket systems, APIs — anything with an MCP server. You enter its address yourself in MCP Server URL

The Topic-to-Topic case is the one worth understanding. Instead of copying documents into one big Topic, you connect a "front desk" Topic to specialised ones and let it ask them.

Practical points:

The calling Topic’s model must be able to call tools. A model without that ability simply ignores the connected tools. This is the most common reason MCP seems not to work.
Write a clear Description. The model reads it to decide when to ask the other Topic. A vague description means the other Topic is asked too seldom, or for the wrong questions.
Each user’s own access counts. The other Topic is asked with the permissions of the user who is chatting. If that user has no access to it, the answer leaves it out and shows a note.
Connections are managed on the MCP Interfaces tab. The Topic must be running to add a connection. No restart is needed.

Browse ready-made tool servers under MCP Tools in the main navigation.

7. What a Topic can ingest

The wizard’s File Types are three checkboxes that you can combine. Files of other types are refused at upload. If you tick none, every file is accepted.

Content Requirement

Content	Requirement
Text files — for example `txt`, `md`, `json`, `yaml`, `csv`, source code	None
Documents — PDF, DOCX, PPTX	A Docling server, to turn them into text
Images — JPG, PNG, GIF, WEBP, BMP, TIFF	An Image Describer Model that can read images. While image files can be uploaded, Describe pictures with a vision model is always on
Structured data — JSON, XML, YAML, TOML, CSV, TSV	Nothing extra for Flatten. Narrate needs a model, for example the Topic’s Language Model

Text files — for example txt, md, json, yaml, csv, source code

None

Documents — PDF, DOCX, PPTX

A Docling server, to turn them into text

Images — JPG, PNG, GIF, WEBP, BMP, TIFF

An Image Describer Model that can read images. While image files can be uploaded, Describe pictures with a vision model is always on

Structured data — JSON, XML, YAML, TOML, CSV, TSV

Nothing extra for Flatten. Narrate needs a model, for example the Topic’s Language Model

8. Typical workflow

Prerequisites — register at least one AI server with a chat model and an embedding model, and a vector database. Add a Docling server if you have PDF or Office files.
Create the Topic — the five-step wizard covers Common Info (name, description, port, access roles, Custom System Prompt), Model Strategy (models, file types, Docling, pictures), Vector DB, Retrieval Strategy (search and routing) and Datasources (chunking). → Creation step by step
Start it — wait for Running.
Upload documents — on the Data Upload tab, and watch the progress. → Configuration and functions
Chat — click the Topic name in the list. → Chat window
Optional — connect other Topics or MCP tools, adjust prompts, grant roles access, or drive the Topic from a program. → REST interface · Management REST API