Ollama (self-hosted) AI Model Provider

Ollama serves open-weight models on machines you control. Point Conferbot at your own endpoint and prompts never leave your infrastructure - while the widget, inbox, handover and analytics stay fully managed. It is the only option here where no customer text reaches a third-party model vendor.

Most guides for building a Ollama chatbot are developer tutorials: an API key, a Python or Node script, and an afternoon spent on the parts that are not the model - the widget, the conversation state, the retrieval, the handover to a human. That is the right approach when the chatbot is the product. It is the wrong shape when you need a working chat widget on a business website this week.

Conferbot treats Ollama as a provider you select, not an integration you build. Paste your own API key, choose Ollama from the provider list, and every AI node in your flow runs on it. The visual builder, widget, inbox, analytics and human handover are unchanged.

Selecting Ollama changes which model generates text. It does not put you in a reduced version of the product. All of this runs on Ollama once selected:

• Free-text answers - the visitor asks in their own words and the model answers from your content rather than a scripted branch.
• Knowledge base retrieval - upload documents or point at your site; the retrieval step is handled for you.
• The AI agent - multi-step conversations that decide what to ask next instead of following a fixed script.
• Intent classification - routing to the right branch, team or queue.
• Summarisation on handover - the summary your agent reads before replying.

The flow, channels, inbox, ticketing, analytics and templates are identical whichever provider you pick. That separation is the point: the model is a setting, not an architecture.

No provider is the right answer for every chatbot. These are the honest trade-offs.

Strengths

• No customer text leaves your infrastructure - the strongest possible data-residency answer
• No per-token cost once the hardware exists
• Complete model choice, including fine-tuned private models
• Works offline and in air-gapped networks

Trade-offs

• You run and monitor the hardware
• Latency and throughput depend entirely on your GPUs
• No managed scaling during traffic spikes

Best suited to

• Healthcare, legal, government and defence
• Anyone whose policy forbids sending customer data to external APIs
• Teams with existing GPU capacity

Model families available

• Any open-weight model Ollama supports (Llama, Mistral, DeepSeek, Gemma and others)

End to end this takes a few minutes, and none of it involves a terminal.

1. Create an API key at https://ollama.com/download
2. Open your chatbot's AI settings in Conferbot
3. Choose Ollama as the provider and paste the key
4. Pick a model and publish - existing flows keep working unchanged

If the chatbot already exists on another provider, selecting Ollama is the only change you make. The flow, content, embed snippet and every deployment channel stay exactly as they are.

What you need first

A Ollama account and an API key. Keys are issued from the Ollama console, and the console will usually show a key once only, so copy it before closing the page.

Ollama serves models from an endpoint you run, so there is no third-party API in the path at all. Conferbot connects to your endpoint and prompts never leave your infrastructure.

How your key is stored

Provider API keys are encrypted at rest and scoped to a single workspace. They are never exposed to the browser, never written to logs in plaintext, and never shared between workspaces or providers. This is the part self-built tutorials most often get wrong, since calling a model API directly from client-side JavaScript puts the key in every visitor's browser.

Where conversations go

Nowhere outside your network. This is the strongest data-residency position available and the reason teams in healthcare, legal, government and defence choose it.

What stays with us

Transcripts, contacts and analytics live in your Conferbot workspace regardless of provider, so changing model does not move or re-key your history.

If the provider is unavailable

Configure a second provider and conversations continue on it instead of failing in front of the visitor. Without a fallback, AI nodes error while the scripted branches and human handover keep working. It costs nothing until it is used.

There are two costs, billed separately by two different parties.

What self-hosting costs

Hardware and the electricity to run it. There is no per-token bill, which changes the economics completely at high volume and makes them worse at low volume.

What Conferbot charges

A conversation-based plan starting free with 600 conversations a month. A conversation is an entire session, not an individual message, so a visitor asking six follow-up questions counts once rather than six times.

Why the split matters

Bring-your-own-key means a cheaper model actually makes your bot cheaper. On platforms that bundle AI into per-message pricing, switching to a cheaper provider saves the vendor money, not you.

Twenty-one providers are wired up, and the right one depends on which constraint binds hardest. This is the short version; the full grid, grouped by jurisdiction and open-weight status, is on the AI model providers page.

On cost

Per-token price only decides the argument at volume. A support bot answering ten thousand questions a month feels it; a lead bot handling two hundred does not. DeepSeek is the cheapest of the mainstream options and has the fullest guide here.

On tone and instruction-following

This is where models differ most and benchmarks help least. Claude is careful and rarely invents; GPT is concise and consistent; Gemini is cheap at the Flash tier with very large context.

On where the data goes

You are already at the strongest position available - nothing leaves your network.

On latency

Groq and Cerebras serve open models on custom silicon fast enough that the widget feels instant, which matters more than model quality on a high-traffic website widget.

On breadth

OpenRouter reaches hundreds of models through one key with automatic failover, at the cost of an extra hop and a routing margin.

Start free: configure two providers on the free plan and run the same flow against both. Twenty of your own questions settle it faster than any table.

You do not have to design the conversation from a blank canvas. The template library holds complete flows you open, edit and publish, and every one works with Ollama because the provider is independent of the flow.

By the job you need done

• Lead generation flows - qualify, capture and route.
• Support and FAQ flows - deflect the repeat questions first.
• Booking and scheduling flows - real availability, confirmed in chat.
• E-commerce flows - order status, returns and product guidance.

By where the bot will live

The same Ollama-powered bot deploys to your website, WhatsApp, Telegram, Discord, Slack, Instagram, Messenger, Microsoft Teams and LINE, with every conversation landing in one inbox. Channel-specific template sets are at website, Telegram and Discord.

Then make it yours

Edit the flow in the visual builder, ground the answers in an AI knowledge base, and set the exit to live chat handover so a person picks up when the bot cannot help.

Most disappointing chatbot results are prompt and grounding problems rather than model problems. These five apply whichever provider you choose.

Ground it before you tune it

Retrieval from an AI knowledge base reduces invention far more reliably than instructions telling the model not to invent. Do this first.

Set an explicit length limit

Chat widgets need two or three sentences, not four paragraphs. Long replies push the input box off a phone screen, and most chat traffic is mobile.

Define the refusal path

Say what to do when the knowledge base does not cover the question. Without that instruction the model attempts an answer; with it you get a clean handover.

Test the failure frame, not the happy path

Ask the bot something it cannot know and watch what happens. That is the frame visitors remember, and the fastest way to find a prompt that is too permissive.

Read the transcripts weekly

Fix the prompt against real failures in the analytics and the inbox, not imagined ones. Twenty real questions beat any benchmark.

Try it: the free plan includes 600 conversations a month, the knowledge base and human handover, with no card and no expiry.

Conferbot supports twenty-one providers, so the honest answer to "is Ollama right for me" is sometimes no. Switching costs one setting.

If you need the safest general-purpose default

OpenAI's GPT models are the most consistent at following varied instructions and the best documented.

If tone and carefulness matter most

Claude is the least likely to invent an answer when the knowledge base does not cover the question.

If EU data residency is the constraint

Mistral is a French company with EU-hosted inference and open weights on several models.

If cost per conversation dominates

DeepSeek is the cheapest of the mainstream options and has a full setup guide here.

If nothing may leave your network

Serve open weights on your own hardware through a self-hosted Ollama endpoint.

The full list, grouped by jurisdiction and open-weight status, is on the AI model providers page.

The free plan includes 600 conversations a month, the visual builder, knowledge base, human handover and every deployment channel, with no credit card and no expiry. Provider choice is not gated by plan, so Ollama works on it from the first conversation.

Still deciding? Configure two providers and run the same flow against both using twenty questions your visitors actually ask. It settles the argument faster than any benchmark table, and leaves you with a fallback configured either way.

From here: pick a flow from the template library, check the limits on the pricing page, compare against other chatbot platforms, or browse every supported provider.