TOLlama (Ollama / Local LLM)

Fonte: source/classes/tollama.prg

TOLlama is FiveWin's client for the Ollama API, enabling interaction with locally-hosted LLMs via Ollama's REST API (default: http://localhost:11434). It supports text chat, streaming responses, and image analysis for vision-capable models. The class uses Harbour's hb_curl library and connects to your local Ollama server, requiring no API keys or internet connectivity. The companion TAgent class provides a data container for tool/agent definitions.

Architecture

flowchart LR subgraph "FiveWin Application" A[TOLlama] end subgraph "libcurl" B[hb_curl
HTTPS] end subgraph "Local Ollama Server" C["localhost:11434/api/chat"] end A --> B B --> C

Key DATA Members

DATA	Type	Default	Description
`cModel`	Character	`"deepseek-r1:14b"`	Ollama model to use
`cPrompt`	Character		Last prompt sent to the API
`cResponse`	Character		Raw JSON response from the API
`cUrl`	Character	`"http://localhost:11434/api/chat"`	Ollama chat API endpoint
`hCurl`	Handle		libcurl easy handle
`nError`	Numeric	`0`	curl error code (0 = `HB_CURLE_OK`)
`nHttpCode`	Numeric	`0`	HTTP response status code
`aAgents`	Array		Array of `TAgent` objects

Methods

Method	Description
`New( cModel )`	Create a TOLlama instance. Sets the model name and initializes the libcurl handle. Default endpoint is `http://localhost:11434/api/chat`.
`Send( cPrompt )`	Send a text prompt to the Ollama chat API. Stores the raw JSON response in `cResponse` and returns it.
`SendStream( cPrompt, bWriteFunction )`	Send a prompt with streaming enabled. The `bWriteFunction` codeblock receives each JSON chunk as it arrives from the server.
`SendImage( cImageFileName, cPrompt )`	Send a local image file with a text prompt for vision analysis. Encodes the image as base64. Supported by vision models like LLaVA.
`GetValue()`	Extract the response text from the JSON in `cResponse`. Returns `message.content` or the error message.
`End()`	Clean up the libcurl handle. Automatically called by the destructor.

TAgent Class

DATA	Type	Description
`cCategory`	Character	Agent category or identifier
`aTools`	Array	Array of tool definitions associated with this agent

TAgent is a data container for defining agents and their associated tool sets. It works alongside TOLlama's aAgents array to organize multiple agent configurations.

Method	Description
`New( cCategory, aTools )`	Create a TAgent instance with a category name and an array of tool definitions.

Example: Local LLM Chat

#include "FiveWin.ch"

function Main()

   local oAI := TOLlama():New( "llama3.2" )

   oAI:Send( "What is the capital of France?" )

   if oAI:nError == 0 .and. oAI:nHttpCode == 200
      MsgInfo( oAI:GetValue() )
   else
      MsgStop( "Error: " + oAI:cResponse )
   endif

   oAI:End()

return nil

Example: Streaming Response

#include "FiveWin.ch"

function Main()

   local oAI := TOLlama():New( "llama3.2" )
   local cBuffer := ""

   oAI:SendStream( "Tell me a joke.", ;
      { |cChunk| cBuffer += cChunk, QOut( cChunk ) } )

   if oAI:nError == 0
      MsgInfo( "Complete response:" + CRLF + cBuffer )
   endif

   oAI:End()

return nil

Example: Image Analysis

#include "FiveWin.ch"

function Main()

   local oAI := TOLlama():New( "llava" )  // vision-capable model

   oAI:SendImage( "C:\photos\photo.jpg", "What objects do you see?" )

   if oAI:nError == 0
      MsgInfo( oAI:GetValue() )
   endif

   oAI:End()

return nil

Notes

TOLlama connects to a locally-running Ollama server. No API key or internet access is required.
The default model is "deepseek-r1:14b" (set in the class), but you should pass the model name you have pulled locally (e.g., "llama3.2", "llava", "mistral").
For image analysis, use a vision-capable model such as "llava" or "bakllava".
GetValue() extracts the message.content field from the Ollama JSON response. On error it returns the error.message field.
The Send() method sets stream: false and temperature: 0.5 in the request. SendStream() sets stream: true.
Requires the hb_curl library. Include #include "hbcurl.ch" for curl constants.