Skip to content

AI Provider Setup

Provider configuration happens via operator environment variables, not the interface

There is no "Settings > AI Provider" screen in the Kamerplanter interface. The AI provider is configured exclusively through environment variables on the Knowledge Service (src/knowledge-service/app/config.py). This is an operator task, not a user setting. Set the variables via Helm values, a .env file, or a Kubernetes secret. This page explains the setup for self-hosters and platform operators. The Ollama installation instructions and hardware recommendations below remain unchanged.

Kamerplanter supports multiple AI providers that can be chosen based on hardware, privacy requirements, and budget. This page explains how to set up each provider and configure it on the Knowledge Service.

Prerequisites

  • Kamerplanter (including the Knowledge Service) is deployed
  • Access to the Knowledge Service's environment variable configuration (Helm values.yaml, .env file, or Kubernetes secret) — operator role

Overview of the Relevant Environment Variables

Variable Description Default
LLM_PROVIDER anthropic, ollama, or openai_compatible ollama
LLM_API_URL Base URL of the provider (for Ollama/OpenAI-compatible) http://ollama:11434
LLM_API_KEY API key (for Anthropic/OpenAI-compatible, if required) empty
LLM_MODEL Model name gemma3:12b
LLM_MAX_TOKENS Maximum response length 2048
LLM_TEMPERATURE Response creativity (0.0–1.0) 0.1

RAM note for the default model

The default gemma3:12b needs significantly more RAM/VRAM than smaller models (see the hardware table below). On smaller hardware, set LLM_MODEL explicitly to a suitable model (e.g. llama3.2:3b or gemma3:4b).

Provider Overview

Provider Type Privacy Cost Recommendation
Ollama Local No data sharing Free Self-hosted
llama.cpp HTTP Server Local No data sharing Free Advanced users
OpenAI API Cloud Transferred to OpenAI (USA) Pay-per-token Best quality
Anthropic Claude API Cloud Transferred to Anthropic (USA) Pay-per-token Best quality
OpenAI-compatible APIs Local or Cloud Depends Variable Advanced users

Recommendation for getting started

If self-hosting: start with Ollama + gemma3:4b. This model runs on most desktop computers from 2020 onwards without a GPU and shares no data externally.

Ollama is a program that runs language models locally on a machine or server. No data leaves the network.

Hardware Requirements

Hardware RAM Recommended Model Response Time (Tip Cards)
Raspberry Pi 5, older NUCs 8 GB llama3.2:3b 15–30 seconds
Desktop/laptop from 2020+ 16 GB gemma3:4b 10–20 seconds
GPU 6–8 GB VRAM (GTX 1060, RX 580) — mistral:7b 2–5 seconds
GPU 12 GB VRAM (RTX 3060) — llama3.1:8b 1–3 seconds
GPU 16 GB VRAM and more — mistral-small:22b 2–5 seconds

Why small models work well

Kamerplanter sends a precise context (current phase, EC/pH/VPD, care history) directly to the model. A 4B model with concrete context produces better plant tips than a 70B model without context.

Installing Ollama

curl -fsSL https://ollama.com/install.sh | sh

The Ollama service starts automatically and is available at http://localhost:11434.

Download the installer from ollama.com/download and open the .dmg file.

After installation, the Ollama icon appears in the menu bar.

Download the installer from ollama.com/download and run it.

Ollama runs as a background service and is available at http://localhost:11434.

docker run -d --name ollama \
  -p 11434:11434 \
  -v ollama_data:/root/.ollama \
  ollama/ollama

With GPU support (NVIDIA):

docker run -d --name ollama \
  --gpus all \
  -p 11434:11434 \
  -v ollama_data:/root/.ollama \
  ollama/ollama

Downloading a Model

Open a terminal and download the recommended model:

# Recommended for most users (16 GB RAM)
ollama pull gemma3:4b

# For machines with less RAM (8 GB)
ollama pull llama3.2:3b

# For GPU users with 8+ GB VRAM
ollama pull mistral:7b

Test Ollama

Verify that Ollama works: 

ollama run gemma3:4b "What temperature does basil need during germination?"

Configuring on the Knowledge Service

Set the following environment variables for the Knowledge Service (e.g. in the Helm values or the .env file) and restart the service:

LLM_PROVIDER=ollama
LLM_API_URL=http://ollama:11434   # or the IP/service name of the Ollama host
LLM_MODEL=gemma3:4b

Ollama on another host

If Ollama runs on a different machine (e.g. a NAS), LLM_API_URL must point to that machine's IP address or DNS name. Port 11434 must be reachable from the Knowledge Service's network.

llama.cpp HTTP Server

llama.cpp is an alternative to Ollama for advanced users who want to use GGUF models directly from the Hugging Face community or other sources.

Starting the Server

# llama.cpp HTTP server (after compilation)
./llama-server \
  --model /path/to/model.gguf \
  --host 0.0.0.0 \
  --port 8080 \
  --ctx-size 4096

Configuring on the Knowledge Service

llama.cpp offers an OpenAI-compatible API, so the openai_compatible provider is used:

LLM_PROVIDER=openai_compatible
LLM_API_URL=http://localhost:8080   # base URL WITHOUT /v1 — the adapter appends /v1/chat/completions itself
LLM_API_KEY=                        # leave empty
LLM_MODEL=local                     # or the name of the loaded GGUF model

OpenAI API

OpenAI provides high-quality cloud models. Plant data is transferred to OpenAI servers in the USA for every request.

Privacy notice

When using the OpenAI API, plant data (phase, measurements, cultivar name, fertilization history) is transferred to OpenAI in the USA. The operator is responsible for disclosing this in the instance's privacy notice.

Creating an API Key

  1. Open platform.openai.com
  2. Sign in (or create an account)
  3. Navigate to API keys
  4. Click Create new secret key
  5. Copy the key (it is only shown once)

Configuring on the Knowledge Service

There is no dedicated openai provider value — OpenAI is reached through openai_compatible with OpenAI's base URL:

LLM_PROVIDER=openai_compatible
LLM_API_URL=https://api.openai.com   # base URL WITHOUT /v1
LLM_API_KEY=sk-...
LLM_MODEL=gpt-4o-mini
Model Strengths Approximate Cost
gpt-4o-mini Fast, affordable, good for simple diagnoses ~$0.001 per request
gpt-4o Best quality, complex reasoning ~$0.01 per request

Anthropic Claude API

Anthropic Claude is an alternative to OpenAI with strong analytical capabilities. Data is also transferred to servers in the USA.

Privacy notice

Analogous to the OpenAI API: plant data is transferred to Anthropic servers in the USA for every request. The operator is responsible for disclosing this in the privacy notice.

Creating an API Key

  1. Open console.anthropic.com
  2. Sign in (or create an account)
  3. Navigate to API Keys
  4. Click Create Key
  5. Copy the key

Configuring on the Knowledge Service

LLM_PROVIDER=anthropic
LLM_API_KEY=sk-ant-...
LLM_MODEL=claude-sonnet-4-20250514   # adapter default if LLM_MODEL is not set
Model Strengths Approximate Cost
claude-haiku-4-5 Very fast, affordable ~$0.001 per request
claude-sonnet-4-6 Precise diagnoses, nuanced responses ~$0.008 per request

OpenAI-Compatible APIs

Many local and cloud services offer an OpenAI-compatible API. These include:

  • LM Studio — GUI application for local models (Windows/macOS/Linux)
  • vLLM — High-performance inference for servers
  • Together AI — Cloud service with open-source models
  • Mistral AI — Cloud service with Mistral models
  • Groq — Very fast cloud inference

Configuring on the Knowledge Service

LLM_PROVIDER=openai_compatible
LLM_API_URL=<base URL of the service without /v1>   # e.g. http://localhost:1234 for LM Studio
LLM_API_KEY=<if required, otherwise empty>
LLM_MODEL=<model name>

LM Studio example

LM Studio starts a local server at http://localhost:1234. LLM_API_URL=http://localhost:1234, LLM_MODEL=lmstudio-community/gemma-3-4b-it-GGUF.

Provider Priority and Fallback

The Knowledge Service currently uses exactly one configured provider (LLM_PROVIDER) — there is no multi-provider configuration with automatic failover between several cloud/local providers.

If no provider is reachable or POST /api/v1/knowledge/ask fails, the rule-based fallback for tip cards applies once that feature is available: the system generates tip cards based on master data and the current phase — without a language model.

Frequently Asked Questions

Can I use different providers for different features?

No, the Knowledge Service currently uses the single configured provider (LLM_PROVIDER) for all AI features. Per-feature provider selection is not supported.

How can I control cloud provider usage costs?

OpenAI and Anthropic provide usage dashboards and budget limits in their control panels.

Ollama won't start or is unreachable — what should I do?

Check: (1) Is the Ollama service running? (systemctl status ollama on Linux). (2) Is Ollama on port 11434? (curl http://localhost:11434). (3) Is the model downloaded? (ollama list). (4) Does LLM_API_URL point to the correct host?

The model responds in a different language — what can I do?

By default, the Knowledge Service sends all requests in German (RAG_PROMPT_LANGUAGE=de). Model behavior also depends on the model itself. If a model still consistently responds in English, a larger model helps (gemma3:4b instead of llama3.2:3b). English-speaking self-hosters can set RAG_PROMPT_LANGUAGE=en to change the default prompt language.

See Also