smartbotic/nodes/ai/ollama-chat.js

/**
 * @node ollama-chat
 * @name Ollama Chat
 * @category ai
 * @version 1.0.0
 * @description Send a prompt to an Ollama model, optionally with an image for vision models, and return the text or parsed JSON response
 * @icon ollama
 */

const configSchema = {
    type: 'object',
    properties: {
        baseUrl: {
            type: 'string',
            title: 'Ollama Base URL',
            description: 'Base URL of the Ollama instance. Use https://ollama.com to call the hosted API directly instead of proxying through a local daemon.',
            default: 'http://localhost:11434'
        },
        authType: {
            type: 'string',
            title: 'Authentication',
            description: 'Hosted ollama.com requires a credential. A local daemon usually does not.',
            enum: ['none', 'credential'],
            enumLabels: ['None', 'Stored credential'],
            default: 'none'
        },
        credentialId: {
            type: 'string',
            title: 'Credential',
            description: 'Stored Bearer credential holding the ollama.com API key',
            // Read as a header, so any header-shaped credential works. An IMAP
            // or database credential has no header to send and would only be
            // an option that fails.
            dynamicOptions: {
                source: 'credentials',
                filter: { type: ['basic', 'bearer', 'api_key', 'oauth2'] }
            },
            default: '',
            showWhen: { field: 'authType', value: 'credential' }
        },
        model: {
            type: 'string',
            title: 'Model',
            description: 'Model tag, for example llama3.2 or a vision model tag',
            default: 'llama3.2'
        },
        systemPrompt: {
            type: 'string',
            title: 'System Prompt',
            description: 'Role and output contract for the model. When Response Format is JSON, mention the word JSON here so the model honours it.',
            format: 'textarea',
            default: ''
        },
        userPrompt: {
            type: 'string',
            title: 'User Prompt',
            description: 'Message sent with the request. Supports {{variable}} interpolation.',
            format: 'textarea',
            default: 'Describe the input.'
        },
        responseFormat: {
            type: 'string',
            title: 'Response Format',
            description: 'How to treat the reply. JSON parses the content and fails the attempt when it is not valid JSON.',
            enum: ['text', 'json'],
            enumLabels: ['Plain text', 'JSON object'],
            default: 'text'
        },
        temperature: {
            type: 'number',
            title: 'Temperature',
            description: 'Sampling temperature. Use 0 for the most repeatable output.',
            default: 0
        },
        imageMode: {
            type: 'string',
            title: 'Image Input',
            description: 'Auto picks up binary or base64 image data from the previous node. Use None for text-only models.',
            enum: ['auto', 'none'],
            enumLabels: ['Auto detect', 'None'],
            default: 'auto'
        },
        imageField: {
            valueKind: 'path',
            type: 'string',
            title: 'Image Field Override',
            description: 'Optional dot path to base64 image data, for example data.file.data',
            default: '',
            showWhen: { field: 'imageMode', value: 'auto' }
        },
        passthroughImage: {
            type: 'boolean',
            title: 'Pass Image Through',
            description: 'Include the base64 image in the output so later nodes can store or reuse it',
            default: false
        },
        retryCount: {
            type: 'number',
            title: 'Retries',
            description: 'Extra attempts after a failed or unparseable response',
            default: 0
        },
        retryDelayMs: {
            type: 'number',
            title: 'Retry Delay (ms)',
            description: 'Pause before the first retry. Doubles on each further attempt.',
            default: 2000
        },
        retryMaxDelayMs: {
            type: 'number',
            title: 'Max Retry Delay (ms)',
            description: 'Upper bound for the exponential backoff',
            default: 30000
        },
        skipOnError: {
            type: 'boolean',
            title: 'Skip On Error',
            description: 'Return success false instead of failing the workflow. Useful inside a loop.',
            default: false
        },
        timeoutMs: {
            type: 'number',
            title: 'Timeout (ms)',
            description: 'Per-attempt request timeout',
            default: 120000
        }
    },
    required: ['baseUrl', 'model']
};

const inputSchema = {
    type: 'object',
    properties: {
        data: {
            type: 'any',
            description: 'Upstream output. Image data is detected here when Image Input is Auto.'
        },
        file: {
            type: 'object',
            description: 'Binary file object with base64 data'
        },
        base64: {
            type: 'string',
            description: 'Raw base64-encoded image'
        },
        url: {
            type: 'string',
            description: 'Image URL, fetched when no binary data is present'
        }
    }
};

const outputSchema = {
    type: 'object',
    properties: {
        success: { type: 'boolean', description: 'False when the call failed and Skip On Error is enabled' },
        error: { type: 'string', description: 'Failure reason when success is false' },
        content: { type: 'string', description: 'Raw text reply from the model' },
        json: { type: 'any', description: 'Parsed reply when Response Format is JSON' },
        model: { type: 'string', description: 'Model that answered' },
        attempts: { type: 'number', description: 'How many attempts were made' },
        hadImage: { type: 'boolean', description: 'Whether an image was sent' },
        imageBase64: { type: 'string', description: 'Base64 image when Pass Image Through is enabled' },
        mimeType: { type: 'string', description: 'Image MIME type when known' },
        sourceUrl: { type: 'string', description: 'Originating image URL when known' }
    }
};

function getPath(root, path) {
    if (!root || !path) {
        return undefined;
    }
    const parts = String(path).split('.');
    let current = root;
    for (let i = 0; i < parts.length; i++) {
        if (current === null || typeof current !== 'object') {
            return undefined;
        }
        current = current[parts[i]];
    }
    return current;
}

// A model asked for JSON that returns something almost-JSON is repaired rather
// than retried. The retry existed for a model having a bad moment, but where
// the fault is deterministic - a model that always omits the opening brace -
// every extra attempt is another paid call for the same malformed answer.
// What was repaired is logged, so a model that has started ignoring the schema
// is visible rather than quietly patched over on every run.
function parseModelJson(text, provider) {
    var repaired = smartbotic.utils.repairJson(text);
    if (!repaired.ok) {
        // Nothing salvageable - let the strict parser raise the real message.
        return JSON.parse(text);
    }
    if (repaired.repairs.length > 0) {
        smartbotic.log.warn(provider + ': the reply was not valid JSON and was repaired (' +
            repaired.repairs.join('; ') + '). The model is not honouring the requested format.');
    }
    return repaired.value;
}

function stripFences(text) {
    const out = String(text || '').trim();
    if (out.indexOf('```') === -1) {
        return out;
    }
    const first = out.indexOf('{');
    const last = out.lastIndexOf('}');
    if (first !== -1 && last !== -1 && last > first) {
        return out.substring(first, last + 1);
    }
    const firstArr = out.indexOf('[');
    const lastArr = out.lastIndexOf(']');
    if (firstArr !== -1 && lastArr !== -1 && lastArr > firstArr) {
        return out.substring(firstArr, lastArr + 1);
    }
    return out;
}

function findImage(input, override) {
    const result = { base64: '', mimeType: '', url: '' };

    if (override) {
        const direct = getPath(input, override);
        if (typeof direct === 'string' && direct.length > 0) {
            result.base64 = direct;
            return result;
        }
    }

    // Collect candidate roots breadth-first: inside a loop, or when another node
    // sits between the fetch and this one, the file object is nested rather than
    // sitting at input or input.data.
    const roots = [];
    const queue = [input];
    let guard = 0;
    while (queue.length > 0 && guard < 64) {
        guard++;
        const node = queue.shift();
        if (!node || typeof node !== 'object') {
            continue;
        }
        roots.push(node);
        const keys = Object.keys(node);
        for (let k = 0; k < keys.length; k++) {
            const child = node[keys[k]];
            if (child && typeof child === 'object' && keys[k] !== 'file') {
                queue.push(child);
            }
        }
    }

    for (let i = 0; i < roots.length; i++) {
        const root = roots[i];
        if (!root || typeof root !== 'object') {
            continue;
        }
        if (!result.base64 && root.file && typeof root.file.data === 'string') {
            result.base64 = root.file.data;
            result.mimeType = root.file.mimeType || '';
        }
        if (!result.base64 && typeof root.base64 === 'string') {
            result.base64 = root.base64;
        }
        if (!result.base64 && typeof root.imageBase64 === 'string') {
            result.base64 = root.imageBase64;
        }
        if (!result.url && typeof root.url === 'string') {
            result.url = root.url;
        }
        if (!result.url && typeof root.sourceUrl === 'string') {
            result.url = root.sourceUrl;
        }
    }
    return result;
}

function buildHeaders(config) {
    const headers = { 'Content-Type': 'application/json' };

    if (config.authType === 'credential' && config.credentialId) {
        const auth = smartbotic.credentials.get(config.credentialId);
        if (!auth.success) {
            throw new Error('Failed to load credential: ' + auth.error);
        }
        headers[auth.headerName] = auth.headerValue;
    }

    return headers;
}

// Statuses no amount of retrying can get past.
//
// 429 is deliberately NOT here: a rate limit is exactly what backing off is
// for. Neither are 500 and 502 - Ollama's cloud returns those transiently, and
// spacing the retries out recovers from them.
const PERMANENT_STATUSES = [
    400,  // bad request - the same payload will be just as bad next time
    401,  // unauthorized - no key, or one the server will not accept
    403,  // refused: no subscription for this model, or the allowance is spent
    404   // no such model
];

function callOllama(config, base64, headers) {
    const messages = [];
    if (config.systemPrompt && String(config.systemPrompt).trim().length > 0) {
        messages.push({ role: 'system', content: String(config.systemPrompt) });
    }

    const userMessage = { role: 'user', content: config.userPrompt || 'Describe the input.' };
    if (base64) {
        userMessage.images = [base64];
    }
    messages.push(userMessage);

    const payload = {
        model: config.model,
        stream: false,
        options: { temperature: Number(config.temperature) || 0 },
        messages: messages
    };

    // Ask the server for JSON, rather than only checking afterwards whether we
    // got any.
    //
    // "JSON object" used to mean nothing more than "parse the reply and fail
    // the attempt if it does not parse". The model was never told, so nothing
    // stopped it emitting almost-JSON - and a reply missing one opening quote
    //
    //   { "title_hu":Tuzfenyes ejszakai olelkezes", "title_en": "..."
    //
    // failed all three attempts and took the workflow down with it. Ollama's
    // format parameter constrains decoding so a malformed reply cannot be
    // produced in the first place; the parse below then only has to deal with
    // an empty or truncated response.
    if (config.responseFormat === 'json') {
        payload.format = 'json';
    }

    const response = smartbotic.http.request({
        method: 'POST',
        url: String(config.baseUrl).replace(/\/+$/, '') + '/api/chat',
        headers: headers,
        body: JSON.stringify(payload),
        timeout: Number(config.timeoutMs) || 120000
    });

    if (response.status < 200 || response.status >= 300) {
        const detail = typeof response.data === 'string' ? response.data : JSON.stringify(response.data);
        const error = new Error('Ollama HTTP ' + response.status + ': ' + detail);
        // Whether trying again could possibly help.
        //
        // Ollama documents 400, 404, 429, 500 and 502; 401 and 403 are not in
        // the documentation but both come back from the cloud endpoint - 401
        // with no key, 403 when the key is fine and the account is refused,
        // which is what an exhausted weekly allowance looks like.
        //
        // Nothing in the permanent set changes because we ask again a few
        // seconds later: the request is malformed, the model does not exist,
        // the key is wrong, or the plan says no. Retrying those only delays
        // the report and spends more of whatever ran out.
        error.permanent = PERMANENT_STATUSES.indexOf(response.status) !== -1;
        error.status = response.status;
        throw error;
    }

    const body = typeof response.data === 'string' ? JSON.parse(response.data) : response.data;
    if (body && body.error) {
        throw new Error('Ollama error: ' + body.error);
    }

    const content = body && body.message ? body.message.content : '';
    if (!content) {
        throw new Error('Ollama returned an empty response');
    }
    return content;
}

module.exports = {
    configSchema,
    inputSchema,
    outputSchema,

    async execute(config, input, context) {
        let image = { base64: '', mimeType: '', url: '' };

        if (config.imageMode !== 'none') {
            image = findImage(input, config.imageField);
            if (!image.base64 && image.url) {
                smartbotic.log.info('ollama-chat: fetching image from ' + image.url);
                const dl = smartbotic.http.request({ method: 'GET', url: image.url, timeout: 60000 });
                if (dl.status < 200 || dl.status >= 300) {
                    throw new Error('Failed to fetch image: HTTP ' + dl.status);
                }
                image.base64 = typeof dl.data === 'string' ? smartbotic.utils.base64Encode(dl.data) : '';
            }
        }

        // Resolved once, outside the retry loop: a missing or broken credential is
        // not transient, so retrying it with backoff only wastes time.
        const headers = buildHeaders(config);

        const wantJson = config.responseFormat === 'json';
        const attempts = 1 + (Number(config.retryCount) > 0 ? Number(config.retryCount) : 0);

        let content = '';
        let parsed = null;
        let lastError = '';
        let used = 0;
        let permanent = false;

        for (let attempt = 1; attempt <= attempts; attempt++) {
            used = attempt;
            try {
                content = callOllama(config, image.base64, headers);
                if (wantJson) {
                    parsed = parseModelJson(stripFences(content), 'Ollama');
                }
                lastError = '';
                break;
            } catch (err) {
                lastError = err && err.message ? err.message : String(err);
                parsed = null;
                if (err && err.permanent === true) {
                    // Said once, and said as what it is. Reporting "failed after
                    // 3 attempts" for a refusal invites the reader to wonder
                    // what was flaky, when nothing was.
                    smartbotic.log.warn('ollama-chat: ' + lastError);
                    permanent = true;
                    break;
                }
                smartbotic.log.warn('ollama-chat: attempt ' + attempt + ' of ' + attempts + ' failed: ' + lastError);
                if (attempt < attempts) {
                    // Exponential backoff. Ollama's cloud tier returns transient 500s far
                    // more often under rapid succession, so spacing retries out recovers
                    // markedly better than hammering at a fixed interval.
                    const base = Number(config.retryDelayMs) || 2000;
                    const cap = Number(config.retryMaxDelayMs) || 30000;
                    let delay = base * Math.pow(2, attempt - 1);
                    if (delay > cap) { delay = cap; }
                    smartbotic.log.info('ollama-chat: backing off ' + delay + 'ms before retry');
                    smartbotic.utils.sleep(delay);
                }
            }
        }

        const passImage = config.passthroughImage === true;

        if (lastError) {
            // A refusal is reported as a refusal. "Failed after 3 attempts"
            // reads as something flaky that might work next time, which sends
            // whoever gets the alert looking for a fault that is not there -
            // the answer to a spent allowance or a wrong key is not to run it
            // again.
            const summary = permanent
                ? 'ollama-chat was refused: ' + lastError
                : 'ollama-chat failed after ' + used + ' attempt(s): ' + lastError;
            if (config.skipOnError !== true) {
                throw new Error(summary);
            }
            smartbotic.log.warn('ollama-chat: skipping - ' + summary);
            return {
                success: false,
                error: lastError,
                content: content,
                json: null,
                model: config.model,
                attempts: used,
                hadImage: image.base64 ? true : false,
                imageBase64: passImage ? image.base64 : '',
                mimeType: image.mimeType,
                sourceUrl: image.url
            };
        }

        return {
            success: true,
            error: '',
            content: content,
            json: parsed,
            model: config.model,
            attempts: used,
            hadImage: image.base64 ? true : false,
            imageBase64: passImage ? image.base64 : '',
            mimeType: image.mimeType,
            sourceUrl: image.url
        };
    }
};