Getting Started — from nothing to a working chatbot

LM Studio is the quickest fully local option: no API key, no cost, no data leaving your machine. Skip this if you intend to use a cloud provider instead (see 2b).

Check: curl localhost:1234/v1/models lists both models.

📖 Detail: Getting started with a local model — installation, quantisation choice, the lms CLI, headless mode

Register the endpoint so its models appear in the Topic wizard: Admin → AI Servers → Add AI Server.

There is a third option: Managed (deployed by this app) starts an Ollama container for you. Use it when you want the platform to own the model server’s lifecycle; it is locked to Ollama.

Check: open the new server, go to the Models tab. Both models are listed. An empty list means the endpoint is not reachable — the Error status in the list means the same thing.

📖 Detail: Connect LM Studio · Connect Claude (Anthropic)

Docling converts PDF, DOCX, PPTX and XLSX into clean text before embedding. Without it, a PDF cannot be indexed at all — and the Topic wizard requires a server as soon as you tick Binary documents among the accepted file types (or allow them as chat attachments).

Go to Admin → Docling Servers. No server is seeded by default, so add one:

The first start pulls a multi-GB image; the status cell names the current phase (Downloading the Docling image (several GB) → Starting the container → Waiting for the Docling server to answer) together with the elapsed time. A managed server has no Host field — its address is filled in when it starts. With Docker you still choose its Port (default 5001; change it only if another program or another managed Docling server already uses that port). With Kubernetes there is no Port field either. 📖 Detail: Docling Servers

Check: the server’s status is RUNNING (managed) or reachable (external).

Every Topic needs a registered vector database. It stores the searchable form of your documents. Register one under Admin → Vector Databases before you create your first Topic. You need an administrator for this. Qdrant, Milvus, Weaviate and CYROCK.DB are supported. Several Topics can share one database — each Topic uses its own collection in it.

If you do not run a vector database yet, start one first, for example a Qdrant container next to the platform.

📖 Detail: Connect a Vector DB — supported types, TLS, the connection test, collection naming, sharing one instance

One look at Admin → LLM Configuration before you build anything, because two of these bite later:

Everything here is a default that each Topic can override: the timeouts and the context memory on its LLM Overrides tab, the system prompt and the no-answer response on its Prompts tab.

📖 Detail: Global LLM configuration & prompt templates

Step 1 — Common Information

Name (letters, digits, hyphens and underscores; spaces become underscores), description, the Custom System Prompt that gives the Topic its character, a free RAG Service Port on the host (8081, 8082, …), and the Access Roles allowed to use it. The prompt is inserted into the global template, so write only what is specific to this Topic — "You are an HR assistant. Answer only from the policy documents."

Step 2 — Models · the step that blocks progress

Tick the File Types you accept first, because it reshapes the rest of the step: ticking Binary documents brings in the Docling section, ticking Images brings in Image Handling and forces "Describe" so ingested images actually get processed. This step also carries the Topic’s Chat Attachments settings, since a binary or image attachment needs the same Docling/vision infrastructure as an upload. Then select the embedding model (it asks for the output dimension — keep the pre-filled native size) and the chat model.

Step 3 — Vector Storage — one dropdown, and you must pick a database (see 2d). The choice is fixed once the Topic is created.

Step 4 — Retrieval Strategy — everything is pre-filled with working defaults (HYBRID, Top-K 3, reranking NONE, minimum similarity 0.6); you can pass straight through.

Step 5 — Data Source — storage mode and the chunking settings covered in 4a. The primary button reads Save instead of Next — that is how you know you are on the last step.

After saving, the Topic is in status CREATED. Press Start: STARTING → RUNNING. The first start downloads the Topic image and can take a few minutes.

Check: the Topic shows RUNNING in the list.

📖 Detail: Creation step by step — every field of all five steps, with the validation rules · Topic overview — what actually runs per Topic

Documents are not embedded whole — they are split into overlapping chunks, and a chunk is the unit that gets retrieved and handed to the model. The settings live in wizard step 5 (and on the Topic’s Configuration tab afterwards):

The two chunking strategies:

SEMANTIC is worth it when a chunk means little on its own — a table row, a numbered clause, a paragraph that says "it must not exceed this value" without naming what it is. Despite the name it does not split on semantic similarity; it improves what each chunk says about itself.

Chunk size and overlap are fixed at creation — changing them would leave old and new embeddings inconsistent, so create a new Topic instead. The strategy can be changed later on the Topic’s Configuration tab: it travels with each upload rather than with the container, so it applies from the next upload on without a restart. Documents already indexed keep the chunking they were given; re-upload them to convert.

The Chunks tab is the most useful diagnostic in the product. It shows exactly what the model can retrieve: garbled text means the conversion failed (usually Docling), chunks cut mid-sentence explain incoherent answers, and a document missing entirely means ingestion failed silently rather than retrieval being at fault.

📖 Detail: Configuration and functions — the upload tab, the three inspection tabs, and what a Database/Server Files mismatch means

Type into the input bar and press Enter or Send. While an answer is being generated the input is disabled and a Stop button appears — or press Esc. A local model taking a minute or more is normal; if it aborts itself, the Chat Timeout is too low.

Under each answer, a Sources: row lists the documents the passages came from, one badge per file. That row is what makes an answer verifiable — if it looks wrong, open the cited file and check. No sources listed means the model answered without retrieval.

Conversations are persistent and per user: + New Chat starts a fresh context, History reopens an earlier one. Deleting a conversation deletes its messages with it.

The 📎 paperclip below the input attaches files to a single question. You can also drag files onto the input field, and pasting more than 300 characters turns the text into an attachment chip instead of flooding the field.

📖 Detail: Chat window — layout, sessions, sources, attachments, error handling, and how to get better answers