ollama_chat

Available in: Self-Managed

Ollama connectors are currently only available on BYOC GCP clusters.

When Redpanda Connect runs a data pipeline with a Ollama processor in it, Redpanda Cloud deploys a GPU-powered instance for the exclusive use of that pipeline. As pricing is based on resource consumption, this can have cost implications.

Generates responses to messages in a chat conversation using the Ollama API and external tools.

# Common configuration fields, showing default values
label: ""
ollama_chat:
  model: llama3.1 # No default (required)
  prompt: "" # No default (optional)
  image: 'root = this.image.decode("base64")' # Decode base64 encoded image. No default (optional)
  response_format: text
  max_tokens: 0 # No default (optional)
  temperature: 0 # No default (optional)
  save_prompt_metadata: false
  history: "" # No default (optional)
  tools: [] # No default (required)
  runner:
    context_size: 0 # No default (optional)
    batch_size: 0 # No default (optional)
  server_address: http://127.0.0.1:11434 # No default (optional)
yml

This processor sends prompts to your chosen Ollama large language model (LLM) and generates text from the responses using the Ollama API and external tools.

By default, the processor starts and runs a locally-installed Ollama server. Alternatively, to use an already running Ollama server, add your server details to the server_address field. You can download and install Ollama from the Ollama website.

For more information, see the Ollama documentation and examples.

Fields

cache_directory

If server_address is not set - the directory to download the Ollama binary and use as a model cache.

Type: string

# Examples:
cache_directory: /opt/cache/connect/ollama
yaml

download_url

If server_address is not set - the URL to download the Ollama binary from. Defaults to the official Ollama GitHub release for this platform.

Type: string

frequency_penalty

Positive values penalize new tokens based on the frequency of their appearance in the text so far. This decreases the model’s likelihood to repeat the same line verbatim.

Type: float

history

Include historical messages in a chat request. You must use a Bloblang query to create an array of objects in the form of [{"role": "", "content":""}] where:

  • role is the sender of the original messages, either system, user, assistant, or tool.

  • content is the text of the original messages.

Type: string

image

An optional image to submit along with the prompt value. The result is a byte array.

Type: string

# Examples:
image: root = this.image.decode("base64") # decode base64 encoded image
yaml

max_tokens

The maximum number of tokens to predict and output. Limiting the amount of output means that requests are processed faster and have a fixed limit on the cost.

Type: int

max_tool_calls

The maximum number of sequential calls you can make to external tools to retrieve additional information to answer a prompt.

Type: int

Default: 3

model

The name of the Ollama LLM to use. For a full list of models, see the Ollama website.

Type: string

# Examples:
model: llama3.1
model: gemma2
model: qwen2
model: phi3
yaml

num_keep

Specify the number of tokens from the initial prompt to retain when the model resets its internal context. By default, this value is set to 4. Use -1 to retain all tokens from the initial prompt.

Type: int

presence_penalty

Positive values penalize new tokens if they have appeared in the text so far. This increases the model’s likelihood to talk about new topics.

Type: float

prompt

The prompt you want to generate a response for. By default, the processor submits the entire payload as a string. This field supports interpolation functions.

Type: string

repeat_penalty

Sets how strongly to penalize repetitions. A higher value, for example 1.5, will penalize repetitions more strongly. A lower value, for example 0.9, will be more lenient.

Type: float

response_format

The format of the response the Ollama model generates. If specifying JSON output, then the prompt should specify that the output should be in JSON as well.

Type: string

Default: text

Options: text, json

runner

Options for the model runner that are used when the model is first loaded into memory.

Type: object

runner.batch_size

The maximum number of requests to process in parallel.

Type: int

runner.context_size

Sets the size of the context window used to generate the next token. Using a larger context window uses more memory and takes longer to process.

Type: int

runner.gpu_layers

This option allows offloading some layers to the GPU for computation. This generally results in increased performance. By default, the runtime decides the number of layers dynamically.

Type: int

runner.threads

Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has. By default, the runtime decides the optimal number of threads.

Type: int

runner.use_mmap

Map the model into memory. This is only support on unix systems and allows loading only the necessary parts of the model as needed.

Type: bool

save_prompt_metadata

Set to true to save the prompt value to a metadata field (@prompt) on the corresponding output message. If you use the system_prompt field, its value is also saved to an @system_prompt metadata field on each output message.

Type: bool

Default: false

seed

Sets the random number seed to use for generation. Setting this to a specific number will make the model generate the same text for the same prompt.

Type: int

# Examples:
seed: 42
yaml

server_address

The address of the Ollama server to use. Leave the field blank and the processor starts and runs a local Ollama server or specify the address of your own local or remote server.

Type: string

# Examples:
server_address: http://127.0.0.1:11434
yaml

stop[]

Sets the stop sequences to use. When this pattern is encountered, the LLM stops generating text and returns the final response.

Type: array

system_prompt

The system prompt to submit to the Ollama LLM. This field supports interpolation functions.

Type: string

temperature

The temperature of the model. Increasing the temperature makes the model answer more creatively.

Type: int

tools[]

The external tools the LLM can invoke, such as functions, APIs, or web browsing. You can build a series of processors that include definitions of these tools, and the specified LLM can choose when to invoke them to help answer a prompt. For more information, see examples.

Type: object

Default: []

tools[].description

A description of this tool, the LLM uses this to decide if the tool should be used.

Type: string

tools[].name

The name of this tool.

Type: string

tools[].parameters

The parameters the LLM needs to provide to invoke this tool.

Type: object

tools[].parameters.properties

The properties for the processor’s input data

Type: object

tools[].parameters.properties.description

A description of this parameter.

Type: string

tools[].parameters.properties.enum[]

Specifies that this parameter is an enum and only these specific values should be used.

Type: array

Default: []

tools[].parameters.properties.type

The type of this parameter.

Type: string

tools[].parameters.required[]

The required parameters for this pipeline.

Type: array

Default: []

tools[].processors[]

The pipeline to execute when the LLM uses this tool.

Type: processor

top_k

Reduces the probability of generating nonsense. A higher value, for example 100, will give more diverse answers. A lower value, for example 10, will be more conservative.

Type: int

top_p

Works together with top-k. A higher value, for example 0.95, will lead to more diverse text. A lower value, for example 0.5, will generate more focused and conservative text.

Type: float

Examples

Use Llava to analyze an image

This example fetches image URLs from stdin and has a multimodal LLM describe the image.

input:
  stdin:
    scanner:
      lines: {}
pipeline:
  processors:
    - http:
        verb: GET
        url: "${!content().string()}"
    - ollama_chat:
        model: llava
        prompt: "Describe the following image"
        image: "root = content()"
output:
  stdout:
    codec: lines
yaml

Use subpipelines as tool calls

This example allows llama3.2 to execute a subpipeline as a tool call to get more data.

input:
  generate:
    count: 1
    mapping: |
      root = "What is the weather like in Chicago?"
pipeline:
  processors:
    - ollama_chat:
        model: llama3.2
        prompt: "${!content().string()}"
        tools:
          - name: GetWeather
            description: "Retrieve the weather for a specific city"
            parameters:
              required: ["city"]
              properties:
                city:
                  type: string
                  description: the city to lookup the weather for
            processors:
              - http:
                  verb: GET
                  url: 'https://wttr.in/${!this.city}?T'
                  headers:
                    # Spoof curl user-ageent to get a plaintext text
                    User-Agent: curl/8.11.1
output:
  stdout: {}
yaml