@drax/ai-back 3.33.0 → 3.35.1

package/src/providers/OllamaAiProvider.ts

@@ -0,0 +1,469 @@
+import {toJSONSchema} from "zod";
+import type {
+    IAIProvider,
+    IPromptContentPart,
+    IPromptMessage,
+    IPromptParams,
+    IPromptResponse,
+    IPromptTool
+} from "../interfaces/IAIProvider";
+import type {AILogService} from "../services/AILogService";
+import type {IAILogBase} from "@drax/ai-share";
+
+type OllamaMessage = {
+    role: "system" | "user" | "assistant" | "tool",
+    content: string,
+    images?: string[],
+    name?: string,
+}
+
+type OllamaToolCall = {
+    function?: {
+        name?: string,
+        arguments?: string | object,
+    }
+}
+
+class OllamaAiProvider implements IAIProvider{
+    protected _baseUrl: string
+    protected _model: string
+    protected _visionModel?: string
+    protected _embeddingModel?: string
+    protected _aiLogService?: AILogService
+
+    constructor(baseUrl: string, model: string, visionModel?: string, embeddingModel?: string, aiLogService?: AILogService) {
+
+        if (!baseUrl) {
+            throw new Error("Ollama AI baseUrl required")
+        }
+        if (!model) {
+            throw new Error("Ollama AI model required")
+        }
+
+        this._baseUrl = baseUrl.replace(/\/+$/, "")
+        this._model = model
+        this._visionModel = visionModel
+        this._embeddingModel = embeddingModel
+        this._aiLogService = aiLogService
+    }
+
+    get model(){
+        if(!this._model){
+            throw new Error("Ollama AI model not found")
+        }
+        return this._model;
+    }
+
+    protected get visionModel(){
+        return this._visionModel
+    }
+
+    protected get embeddingModel(){
+        return this._embeddingModel ?? this.model
+    }
+
+    protected async post<T>(path: string, body: object): Promise<T> {
+        const response = await fetch(`${this._baseUrl}${path}`, {
+            method: "POST",
+            headers: {
+                "Content-Type": "application/json",
+            },
+            body: JSON.stringify(body),
+        })
+
+        if(!response.ok){
+            const errorText = await response.text()
+            throw new Error(`Ollama AI request failed (${response.status}): ${errorText}`)
+        }
+
+        return await response.json() as T
+    }
+
+    protected async buildUserMessage(input: IPromptParams): Promise<OllamaMessage> {
+        if(input.userContent && input.userContent.length > 0){
+            return await this.mapContentPartsToMessage(input.userContent)
+        }
+
+        if(input.userImages && input.userImages.length > 0){
+            return {
+                role: "user",
+                content: input.userInput ?? "",
+                images: await Promise.all(input.userImages.map(image => this.imageUrlToBase64(image.url))),
+            }
+        }
+
+        return {
+            role: "user",
+            content: input.userInput ?? "",
+        }
+    }
+
+    protected async mapContentPartsToMessage(content: IPromptContentPart[], role: "user" | "assistant" | "system" = "user"): Promise<OllamaMessage> {
+        const text: string[] = []
+        const images: string[] = []
+
+        for(const part of content){
+            if(part.type === "text"){
+                text.push(part.text)
+                continue
+            }
+
+            images.push(await this.imageUrlToBase64(part.imageUrl))
+        }
+
+        return {
+            role,
+            content: text.join("\n"),
+            ...(images.length > 0 ? {images} : {}),
+        }
+    }
+
+    protected async imageUrlToBase64(url: string): Promise<string> {
+        const dataUrlMatch = url.match(/^data:[^;,]+;base64,(.+)$/)
+
+        if(dataUrlMatch){
+            return dataUrlMatch[1]
+        }
+
+        const response = await fetch(url)
+
+        if(!response.ok){
+            throw new Error(`Ollama AI image request failed (${response.status}): ${url}`)
+        }
+
+        const buffer = Buffer.from(await response.arrayBuffer())
+        return buffer.toString("base64")
+    }
+
+    protected async mapHistory(history: IPromptMessage[] = []): Promise<OllamaMessage[]>{
+        const messages: OllamaMessage[] = []
+
+        for(const message of history){
+            if(typeof message.content === "string"){
+                messages.push({
+                    role: message.role,
+                    content: message.content,
+                })
+                continue
+            }
+
+            messages.push(await this.mapContentPartsToMessage(message.content, message.role))
+        }
+
+        return messages
+    }
+
+    protected hasImageInput(input: IPromptParams){
+        if(input.userImages && input.userImages.length > 0){
+            return true
+        }
+
+        if(input.userContent?.some(part => part.type === 'image')){
+            return true
+        }
+
+        return input.history?.some(message =>
+            Array.isArray(message.content) && message.content.some(part => part.type === 'image')
+        ) ?? false
+    }
+
+    protected serializePromptInput(input: IPromptParams, systemPrompt: string){
+        return JSON.stringify({
+            systemPrompt,
+            history: input.history,
+            userInput: input.userInput,
+            userContent: input.userContent,
+            memory: input.memory,
+            knowledgeBase: input.knowledgeBase,
+            tools: input.tools?.map(tool => ({
+                name: tool.name,
+                description: tool.description,
+                parameters: tool.parameters,
+            })),
+        })
+    }
+
+    protected serializePromptOutput(output: unknown){
+        if (typeof output === "string") {
+            return output
+        }
+
+        if (output === null || output === undefined) {
+            return undefined
+        }
+
+        return JSON.stringify(output)
+    }
+
+    protected buildLogPayload(input: IPromptParams, params: {
+        model: string,
+        systemPrompt: string,
+        startedAt: Date,
+        endedAt?: Date,
+        inputTokens?: number,
+        outputTokens?: number,
+        tokens?: number,
+        output?: unknown,
+        success: boolean,
+        errorMessage?: string,
+    }): IAILogBase {
+        return {
+            provider: "ollamaai",
+            model: params.model,
+            operationTitle: input.operationTitle,
+            operationGroup: input.operationGroup,
+            ip: input.ip,
+            userAgent: input.userAgent,
+            input: this.serializePromptInput(input, params.systemPrompt),
+            inputImages: input.userImages?.map(image => ({
+                url: image.url,
+            })) ?? input.userContent
+                ?.filter(part => part.type === "image")
+                .map(part => ({
+                    url: part.imageUrl,
+                })),
+            inputFiles: input.inputFiles,
+            inputTokens: params.inputTokens,
+            outputTokens: params.outputTokens,
+            tokens: params.tokens,
+            startedAt: params.startedAt,
+            endedAt: params.endedAt,
+            responseTime: params.endedAt ? `${params.endedAt.getTime() - params.startedAt.getTime()}ms` : undefined,
+            output: this.serializePromptOutput(params.output),
+            success: params.success,
+            errorMessage: params.errorMessage,
+            tenant: input.tenant,
+            user: input.user,
+        }
+    }
+
+    protected async registerPromptLog(input: IPromptParams, params: {
+        model: string,
+        systemPrompt: string,
+        startedAt: Date,
+        endedAt?: Date,
+        inputTokens?: number,
+        outputTokens?: number,
+        tokens?: number,
+        output?: unknown,
+        success: boolean,
+        errorMessage?: string,
+    }){
+        if(!this._aiLogService){
+            return
+        }
+
+        try{
+            await this._aiLogService.create(this.buildLogPayload(input, params))
+        }catch(e: any){
+            console.error("Error registerPromptLog", {
+                name: e?.name,
+                message: e?.message,
+                stack: e?.stack,
+            })
+        }
+    }
+
+    async generateEmbedding({text, model}: {text:string, model?:string }): Promise<number[]> {
+        const response = await this.post<any>("/api/embed", {
+            model: model ?? this.embeddingModel,
+            input: text,
+        });
+
+        return response.embeddings?.[0] ?? response.embedding ?? [];
+    }
+
+    protected mapTools(tools: IPromptTool[] = []){
+        return tools.map(tool => ({
+            type: "function",
+            function: {
+                name: tool.name,
+                description: tool.description,
+                parameters: tool.parameters ?? {
+                    type: "object",
+                    properties: {},
+                    additionalProperties: false,
+                },
+            },
+        }))
+    }
+
+    protected normalizeResponseFormat(input: IPromptParams){
+        if(input.zodSchema){
+            return toJSONSchema(input.zodSchema, {
+                target: "draft-7",
+            })
+        }
+
+        if(!input.jsonSchema){
+            return undefined
+        }
+
+        const jsonSchema: any = input.jsonSchema
+
+        if(jsonSchema.type === "json_schema" && jsonSchema.json_schema?.schema){
+            return jsonSchema.json_schema.schema
+        }
+
+        return jsonSchema
+    }
+
+    protected parseToolArguments(args: string | object | undefined){
+        if(!args){
+            return {}
+        }
+
+        if(typeof args === "object"){
+            return args
+        }
+
+        try{
+            return JSON.parse(args)
+        }catch(e){
+            throw new Error(`Invalid tool arguments: ${args}`)
+        }
+    }
+
+    protected serializeToolOutput(output: unknown){
+        if(typeof output === "string"){
+            return output
+        }
+
+        if(output === undefined){
+            return ""
+        }
+
+        return JSON.stringify(output)
+    }
+
+    protected async buildToolMessages(toolCalls: OllamaToolCall[] = [], tools: IPromptTool[] = []){
+        const toolMessages: OllamaMessage[] = []
+
+        for(const toolCall of toolCalls){
+            const toolName = toolCall.function?.name
+            const tool = tools.find(t => t.name === toolName)
+
+            if(!tool){
+                throw new Error(`Tool not found: ${toolName}`)
+            }
+
+            const args = this.parseToolArguments(toolCall.function?.arguments)
+            const output = await tool.execute(args)
+
+            toolMessages.push({
+                role: "tool",
+                name: toolName,
+                content: this.serializeToolOutput(output),
+            })
+        }
+
+        return toolMessages
+    }
+
+    async prompt(input: IPromptParams): Promise<IPromptResponse> {
+
+        if(!input.systemPrompt){
+            throw new Error("systemPrompt required")
+        }
+
+        let systemPrompt = input.systemPrompt
+
+        if(input.memory && input.memory.length > 0){
+            systemPrompt += `\n\n ${input.memoryHeader ?? '[MEMORIA]'}\n ${input.memory.map(m => `${m.key}: ${m.value}`).join('\n')}`
+        }
+
+        if(input.knowledgeBase && input.knowledgeBase.length > 0){
+            systemPrompt += `\n\n${input.knowledgeBaseHeader ?? '[BASE DE CONOCIMIENTO]'}\n ${input.knowledgeBase.join('\n')}`
+        }
+
+        const model = input.model ?? (this.hasImageInput(input) ? this.visionModel ?? this.model : this.model)
+        const startedAt = new Date()
+        const startTime = performance.now()
+        let tokens = 0
+        let inputTokens = 0
+        let outputTokens = 0
+
+        try {
+            const messages: OllamaMessage[] = [
+                {role: 'system', content: systemPrompt},
+                ...await this.mapHistory(input.history),
+                await this.buildUserMessage(input),
+            ]
+            const tools = input.tools ?? []
+            const maxIterations = input.toolMaxIterations ?? 5
+            const responseFormat = this.normalizeResponseFormat(input)
+            let output: any
+
+            for(let iteration = 0; iteration < maxIterations; iteration++){
+                const response = await this.post<any>("/api/chat", {
+                    model,
+                    messages,
+                    stream: false,
+                    ...(responseFormat ? {format: responseFormat} : {}),
+                    ...(tools.length > 0 ? {tools: this.mapTools(tools)} : {}),
+                });
+
+                inputTokens += response.prompt_eval_count ?? 0
+                outputTokens += response.eval_count ?? 0
+                tokens += (response.prompt_eval_count ?? 0) + (response.eval_count ?? 0)
+
+                const message = response.message ?? {}
+                const toolCalls = message.tool_calls ?? []
+
+                if(toolCalls.length === 0){
+                    output = message.content ?? response.response ?? ""
+                    break
+                }
+
+                messages.push(message)
+                messages.push(...await this.buildToolMessages(toolCalls, tools))
+            }
+
+            if(output === undefined){
+                throw new Error(`Tool max iterations reached: ${maxIterations}`)
+            }
+
+            const endTime = performance.now()
+            const time = endTime - startTime
+            const endedAt = new Date()
+
+            await this.registerPromptLog(input, {
+                model,
+                systemPrompt,
+                startedAt,
+                endedAt,
+                inputTokens,
+                outputTokens,
+                tokens,
+                output,
+                success: true,
+            })
+
+            return {
+                output,
+                tokens,
+                inputTokens,
+                outputTokens,
+                time
+            }
+        } catch (e: any) {
+            const endedAt = new Date()
+
+            await this.registerPromptLog(input, {
+                model,
+                systemPrompt,
+                startedAt,
+                endedAt,
+                success: false,
+                errorMessage: e?.message,
+            })
+
+            throw e
+        }
+    }
+
+}
+
+
+export default OllamaAiProvider
+export {OllamaAiProvider}