@stravigor/saina 0.4.7

package/src/helpers.ts

@@ -0,0 +1,756 @@
+import SainaManager from './saina_manager.ts'
+import { Agent } from './agent.ts'
+import { Workflow } from './workflow.ts'
+import { zodToJsonSchema } from './utils/schema.ts'
+import type {
+  AIProvider,
+  CompletionRequest,
+  CompletionResponse,
+  StreamChunk,
+  Message,
+  ToolCall,
+  ToolCallRecord,
+  ToolDefinition,
+  AgentResult,
+  AgentEvent,
+  Usage,
+  JsonSchema,
+  SerializedThread,
+} from './types.ts'
+
+// ── Shared tool executor ─────────────────────────────────────────────────────
+
+/** Execute a single tool call, returning the result and the tool message. */
+async function executeTool(
+  tools: ToolDefinition[] | undefined,
+  toolCall: ToolCall
+): Promise<{ result: unknown; message: Message }> {
+  const toolDef = tools?.find(t => t.name === toolCall.name)
+  let result: unknown
+
+  if (!toolDef) {
+    result = `Error: Tool "${toolCall.name}" not found`
+  } else {
+    try {
+      result = await toolDef.execute(toolCall.arguments)
+    } catch (err) {
+      result = `Error: ${err instanceof Error ? err.message : String(err)}`
+    }
+  }
+
+  return {
+    result,
+    message: {
+      role: 'tool',
+      toolCallId: toolCall.id,
+      content: typeof result === 'string' ? result : JSON.stringify(result),
+    },
+  }
+}
+
+// ── Helper Options ───────────────────────────────────────────────────────────
+
+export interface ChatOptions {
+  provider?: string
+  model?: string
+  system?: string
+  maxTokens?: number
+  temperature?: number
+}
+
+export interface GenerateOptions<T = any> {
+  prompt: string
+  schema: any
+  provider?: string
+  model?: string
+  system?: string
+  maxTokens?: number
+  temperature?: number
+}
+
+export interface GenerateResult<T = any> {
+  data: T
+  text: string
+  usage: Usage
+}
+
+export interface EmbedOptions {
+  provider?: string
+  model?: string
+}
+
+// ── saina Helper Object ─────────────────────────────────────────────────────
+
+export const saina = {
+  /**
+   * One-shot chat completion. Returns the text response.
+   *
+   * @example
+   * const answer = await saina.chat('What is the capital of France?')
+   * const answer = await saina.chat('Explain X', { provider: 'openai', model: 'gpt-4o-mini' })
+   */
+  async chat(prompt: string, options: ChatOptions = {}): Promise<string> {
+    const config = SainaManager.config
+    const providerName = options.provider ?? config.default
+
+    const response = await SainaManager.complete(providerName, {
+      model:
+        (options.model ?? SainaManager.provider(providerName).name === 'anthropic')
+          ? (SainaManager.config.providers[providerName]?.model ?? config.default)
+          : (SainaManager.config.providers[providerName]?.model ?? ''),
+      messages: [{ role: 'user', content: prompt }],
+      system: options.system,
+      maxTokens: options.maxTokens ?? config.maxTokens,
+      temperature: options.temperature ?? config.temperature,
+    })
+
+    return response.content
+  },
+
+  /**
+   * One-shot streaming completion.
+   *
+   * @example
+   * for await (const chunk of saina.stream('Write a poem')) {
+   *   if (chunk.type === 'text') process.stdout.write(chunk.text!)
+   * }
+   */
+  async *stream(prompt: string, options: ChatOptions = {}): AsyncIterable<StreamChunk> {
+    const config = SainaManager.config
+    const providerName = options.provider ?? config.default
+    const provider = SainaManager.provider(providerName)
+    const providerConfig = config.providers[providerName]
+
+    yield* provider.stream({
+      model: options.model ?? providerConfig?.model ?? '',
+      messages: [{ role: 'user', content: prompt }],
+      system: options.system,
+      maxTokens: options.maxTokens ?? config.maxTokens,
+      temperature: options.temperature ?? config.temperature,
+    })
+  },
+
+  /**
+   * Structured output completion. Returns typed data validated against the schema.
+   *
+   * @example
+   * const { data } = await saina.generate({
+   *   prompt: 'Extract: "John is 30"',
+   *   schema: z.object({ name: z.string(), age: z.number() }),
+   * })
+   * // data.name === 'John', data.age === 30
+   */
+  async generate<T>(options: GenerateOptions<T>): Promise<GenerateResult<T>> {
+    const config = SainaManager.config
+    const providerName = options.provider ?? config.default
+    const providerConfig = config.providers[providerName]
+    const jsonSchema = zodToJsonSchema(options.schema)
+
+    const response = await SainaManager.complete(providerName, {
+      model: options.model ?? providerConfig?.model ?? '',
+      messages: [{ role: 'user', content: options.prompt }],
+      system: options.system,
+      schema: jsonSchema,
+      maxTokens: options.maxTokens ?? config.maxTokens,
+      temperature: options.temperature ?? config.temperature,
+    })
+
+    const parsed = JSON.parse(response.content)
+    const data = options.schema?.parse ? options.schema.parse(parsed) : parsed
+
+    return {
+      data,
+      text: response.content,
+      usage: response.usage,
+    }
+  },
+
+  /**
+   * Generate embeddings for the given text(s).
+   *
+   * @example
+   * const vectors = await saina.embed('Hello world', { provider: 'openai' })
+   */
+  async embed(input: string | string[], options: EmbedOptions = {}): Promise<number[][]> {
+    const providerName = options.provider ?? SainaManager.config.default
+    const provider = SainaManager.provider(providerName)
+
+    if (!provider.embed) {
+      throw new Error(`Provider "${providerName}" does not support embeddings.`)
+    }
+
+    const result = await provider.embed(input, options.model)
+    return result.embeddings
+  },
+
+  /** Create a fluent agent runner. */
+  agent<T extends Agent>(AgentClass: new () => T): AgentRunner<T> {
+    return new AgentRunner(AgentClass)
+  },
+
+  /** Create a multi-turn conversation thread. */
+  thread(AgentClass?: new () => Agent): Thread {
+    return new Thread(AgentClass)
+  },
+
+  /** Create a multi-agent workflow. */
+  workflow(name: string): Workflow {
+    return new Workflow(name)
+  },
+}
+
+// ── AgentRunner ──────────────────────────────────────────────────────────────
+
+/**
+ * Fluent builder for running an agent. Handles the tool-use loop,
+ * structured output parsing, and lifecycle hooks.
+ *
+ * @example
+ * const result = await saina.agent(SupportAgent)
+ *   .input('Where is my order #12345?')
+ *   .with({ orderId: '12345' })
+ *   .run()
+ */
+export class AgentRunner<T extends Agent = Agent> {
+  private _input = ''
+  private _context: Record<string, unknown> = {}
+  private _provider?: string
+  private _model?: string
+
+  constructor(private AgentClass: new () => T) {}
+
+  /** Set the user input / prompt for the agent. */
+  input(text: string): this {
+    this._input = text
+    return this
+  }
+
+  /** Add context variables. Available as `{{key}}` in agent instructions. */
+  with(context: Record<string, unknown>): this {
+    Object.assign(this._context, context)
+    return this
+  }
+
+  /** Override the provider (and optionally model) for this run. */
+  using(provider: string, model?: string): this {
+    this._provider = provider
+    if (model) this._model = model
+    return this
+  }
+
+  /** Run the agent to completion. */
+  async run(): Promise<AgentResult> {
+    const agent = new this.AgentClass()
+    const config = SainaManager.config
+
+    const providerName = this._provider ?? agent.provider ?? config.default
+    const providerConfig = config.providers[providerName]
+    const model = this._model ?? agent.model ?? providerConfig?.model ?? ''
+    const maxIterations = agent.maxIterations ?? config.maxIterations
+    const maxTokens = agent.maxTokens ?? config.maxTokens
+    const temperature = agent.temperature ?? config.temperature
+
+    try {
+      await agent.onStart?.(this._input, this._context)
+    } catch (err) {
+      await agent.onError?.(err instanceof Error ? err : new Error(String(err)))
+      throw err
+    }
+
+    // Build system prompt with context interpolation
+    let system: string | undefined = agent.instructions || undefined
+    if (system) {
+      for (const [key, value] of Object.entries(this._context)) {
+        system = system.replaceAll(`{{${key}}}`, String(value))
+      }
+    }
+
+    // Prepare structured output schema
+    let schema: JsonSchema | undefined
+    if (agent.output) {
+      schema = zodToJsonSchema(agent.output)
+    }
+
+    const messages: Message[] = [{ role: 'user', content: this._input }]
+    const allToolCalls: ToolCallRecord[] = []
+    const totalUsage: Usage = { inputTokens: 0, outputTokens: 0, totalTokens: 0 }
+    let iterations = 0
+
+    // Tool loop
+    while (iterations < maxIterations) {
+      iterations++
+
+      const request: CompletionRequest = {
+        model,
+        messages: [...messages],
+        system,
+        maxTokens,
+        temperature,
+      }
+
+      // Only send tools if the agent has them
+      if (agent.tools?.length) {
+        request.tools = agent.tools
+      }
+
+      // Only send schema when we're not mid-tool-loop (avoid conflicting constraints)
+      if (schema && (!agent.tools?.length || iterations > 1)) {
+        request.schema = schema
+      }
+
+      let response: CompletionResponse
+      try {
+        response = await SainaManager.complete(providerName, request)
+      } catch (err) {
+        await agent.onError?.(err instanceof Error ? err : new Error(String(err)))
+        throw err
+      }
+
+      // Accumulate usage
+      totalUsage.inputTokens += response.usage.inputTokens
+      totalUsage.outputTokens += response.usage.outputTokens
+      totalUsage.totalTokens += response.usage.totalTokens
+
+      // Append assistant message
+      messages.push({
+        role: 'assistant',
+        content: response.content,
+        toolCalls: response.toolCalls.length > 0 ? response.toolCalls : undefined,
+      })
+
+      // If no tool calls, we're done
+      if (response.stopReason !== 'tool_use' || response.toolCalls.length === 0) {
+        let data: any = response.content
+        if (agent.output && response.content) {
+          try {
+            const parsed = JSON.parse(response.content)
+            data = agent.output.parse ? agent.output.parse(parsed) : parsed
+          } catch {
+            data = response.content
+          }
+        }
+
+        const result: AgentResult = {
+          data,
+          text: response.content,
+          toolCalls: allToolCalls,
+          messages,
+          usage: totalUsage,
+          iterations,
+        }
+
+        await agent.onComplete?.(result)
+        return result
+      }
+
+      // Execute tool calls
+      await this.executeTools(agent, response.toolCalls, messages, allToolCalls)
+    }
+
+    // Max iterations reached — return what we have
+    const lastAssistant = [...messages].reverse().find(m => m.role === 'assistant')
+    const text = typeof lastAssistant?.content === 'string' ? lastAssistant.content : ''
+
+    const result: AgentResult = {
+      data: null,
+      text,
+      toolCalls: allToolCalls,
+      messages,
+      usage: totalUsage,
+      iterations,
+    }
+
+    await agent.onComplete?.(result)
+    return result
+  }
+
+  /** Run the agent with streaming, yielding events for each text chunk and tool execution. */
+  async *stream(): AsyncIterable<AgentEvent> {
+    const agent = new this.AgentClass()
+    const config = SainaManager.config
+
+    const providerName = this._provider ?? agent.provider ?? config.default
+    const providerConfig = config.providers[providerName]
+    const model = this._model ?? agent.model ?? providerConfig?.model ?? ''
+    const maxIterations = agent.maxIterations ?? config.maxIterations
+    const maxTokens = agent.maxTokens ?? config.maxTokens
+    const temperature = agent.temperature ?? config.temperature
+    const provider = SainaManager.provider(providerName)
+
+    await agent.onStart?.(this._input, this._context)
+
+    let system: string | undefined = agent.instructions || undefined
+    if (system) {
+      for (const [key, value] of Object.entries(this._context)) {
+        system = system.replaceAll(`{{${key}}}`, String(value))
+      }
+    }
+
+    let schema: JsonSchema | undefined
+    if (agent.output) {
+      schema = zodToJsonSchema(agent.output)
+    }
+
+    const messages: Message[] = [{ role: 'user', content: this._input }]
+    const allToolCalls: ToolCallRecord[] = []
+    const totalUsage: Usage = { inputTokens: 0, outputTokens: 0, totalTokens: 0 }
+    let iterations = 0
+
+    while (iterations < maxIterations) {
+      iterations++
+
+      if (iterations > 1) {
+        yield { type: 'iteration', iteration: iterations }
+      }
+
+      const request: CompletionRequest = {
+        model,
+        messages: [...messages],
+        system,
+        maxTokens,
+        temperature,
+      }
+
+      if (agent.tools?.length) request.tools = agent.tools
+      if (schema && (!agent.tools?.length || iterations > 1)) request.schema = schema
+
+      // Stream the response
+      let fullText = ''
+      const pendingToolCalls: Map<number, { id: string; name: string; args: string }> = new Map()
+
+      for await (const chunk of provider.stream(request)) {
+        if (chunk.type === 'text' && chunk.text) {
+          fullText += chunk.text
+          yield { type: 'text', text: chunk.text }
+        } else if (chunk.type === 'tool_start' && chunk.toolCall) {
+          pendingToolCalls.set(chunk.toolIndex ?? 0, {
+            id: chunk.toolCall.id ?? '',
+            name: chunk.toolCall.name ?? '',
+            args: '',
+          })
+        } else if (chunk.type === 'tool_delta' && chunk.text) {
+          const pending = pendingToolCalls.get(chunk.toolIndex ?? 0)
+          if (pending) pending.args += chunk.text
+        } else if (chunk.type === 'usage' && chunk.usage) {
+          totalUsage.inputTokens += chunk.usage.inputTokens
+          totalUsage.outputTokens += chunk.usage.outputTokens
+          totalUsage.totalTokens += chunk.usage.totalTokens
+        }
+      }
+
+      // Build tool calls from accumulated stream data
+      const toolCalls: ToolCall[] = []
+      for (const [, pending] of pendingToolCalls) {
+        let args: Record<string, unknown> = {}
+        try {
+          args = JSON.parse(pending.args)
+        } catch {
+          args = pending.args ? { _raw: pending.args } : {}
+        }
+        toolCalls.push({ id: pending.id, name: pending.name, arguments: args })
+      }
+
+      // Append assistant message
+      messages.push({
+        role: 'assistant',
+        content: fullText,
+        toolCalls: toolCalls.length > 0 ? toolCalls : undefined,
+      })
+
+      // If no tool calls, done
+      if (toolCalls.length === 0) {
+        let data: any = fullText
+        if (agent.output && fullText) {
+          try {
+            const parsed = JSON.parse(fullText)
+            data = agent.output.parse ? agent.output.parse(parsed) : parsed
+          } catch {
+            data = fullText
+          }
+        }
+
+        const result: AgentResult = {
+          data,
+          text: fullText,
+          toolCalls: allToolCalls,
+          messages,
+          usage: totalUsage,
+          iterations,
+        }
+
+        await agent.onComplete?.(result)
+        yield { type: 'done', result }
+        return
+      }
+
+      // Execute tools and yield events
+      for (const toolCall of toolCalls) {
+        await agent.onToolCall?.(toolCall)
+
+        const start = performance.now()
+        const { result: toolResult, message } = await executeTool(agent.tools, toolCall)
+        const duration = performance.now() - start
+
+        const record: ToolCallRecord = {
+          name: toolCall.name,
+          arguments: toolCall.arguments,
+          result: toolResult,
+          duration,
+        }
+        allToolCalls.push(record)
+        await agent.onToolResult?.(record)
+
+        yield { type: 'tool_result', toolCall: record }
+
+        messages.push(message)
+      }
+    }
+
+    // Max iterations
+    const lastAssistant = [...messages].reverse().find(m => m.role === 'assistant')
+    const text = typeof lastAssistant?.content === 'string' ? lastAssistant.content : ''
+
+    const result: AgentResult = {
+      data: null,
+      text,
+      toolCalls: allToolCalls,
+      messages,
+      usage: totalUsage,
+      iterations,
+    }
+
+    await agent.onComplete?.(result)
+    yield { type: 'done', result }
+  }
+
+  // ── Private ────────────────────────────────────────────────────────────────
+
+  private async executeTools(
+    agent: Agent,
+    toolCalls: ToolCall[],
+    messages: Message[],
+    allToolCalls: ToolCallRecord[]
+  ): Promise<void> {
+    for (const toolCall of toolCalls) {
+      await agent.onToolCall?.(toolCall)
+
+      const start = performance.now()
+      const { result, message } = await executeTool(agent.tools, toolCall)
+      const duration = performance.now() - start
+
+      const record: ToolCallRecord = {
+        name: toolCall.name,
+        arguments: toolCall.arguments,
+        result,
+        duration,
+      }
+      allToolCalls.push(record)
+      await agent.onToolResult?.(record)
+
+      messages.push(message)
+    }
+  }
+}
+
+// ── Thread ────────────────────────────────────────────────────────────────────
+
+/**
+ * Multi-turn conversation thread with optional agent configuration.
+ *
+ * @example
+ * const thread = saina.thread()
+ * thread.system('You are a helpful assistant.')
+ * const r1 = await thread.send('My name is Alice')
+ * const r2 = await thread.send('What is my name?')
+ *
+ * // With agent
+ * const thread = saina.thread(SupportAgent)
+ * const r1 = await thread.send('I need help with order #123')
+ */
+export class Thread {
+  private messages: Message[] = []
+  private _provider?: string
+  private _model?: string
+  private _system?: string
+  private _tools?: ToolDefinition[]
+  private _maxTokens?: number
+  private _temperature?: number
+
+  constructor(AgentClass?: new () => Agent) {
+    if (AgentClass) {
+      const agent = new AgentClass()
+      this._provider = agent.provider
+      this._model = agent.model
+      this._system = agent.instructions || undefined
+      this._tools = agent.tools
+      this._maxTokens = agent.maxTokens
+      this._temperature = agent.temperature
+    }
+  }
+
+  /** Set or override the system prompt. */
+  system(prompt: string): this {
+    this._system = prompt
+    return this
+  }
+
+  /** Override the provider (and optionally model). */
+  using(provider: string, model?: string): this {
+    this._provider = provider
+    if (model) this._model = model
+    return this
+  }
+
+  /** Set tools available in this thread. */
+  tools(tools: ToolDefinition[]): this {
+    this._tools = tools
+    return this
+  }
+
+  /** Send a message and get the assistant's response. Handles tool calls automatically. */
+  async send(message: string): Promise<string> {
+    const config = SainaManager.config
+    const providerName = this._provider ?? config.default
+    const providerConfig = config.providers[providerName]
+
+    this.messages.push({ role: 'user', content: message })
+
+    const maxIterations = 10
+    let iterations = 0
+
+    while (iterations < maxIterations) {
+      iterations++
+
+      const request: CompletionRequest = {
+        model: this._model ?? providerConfig?.model ?? '',
+        messages: [...this.messages],
+        system: this._system,
+        maxTokens: this._maxTokens ?? config.maxTokens,
+        temperature: this._temperature ?? config.temperature,
+      }
+
+      if (this._tools?.length) request.tools = this._tools
+
+      const response = await SainaManager.complete(providerName, request)
+
+      this.messages.push({
+        role: 'assistant',
+        content: response.content,
+        toolCalls: response.toolCalls.length > 0 ? response.toolCalls : undefined,
+      })
+
+      // If no tool calls, return
+      if (response.stopReason !== 'tool_use' || response.toolCalls.length === 0) {
+        return response.content
+      }
+
+      // Execute tools
+      for (const toolCall of response.toolCalls) {
+        const { message } = await executeTool(this._tools, toolCall)
+        this.messages.push(message)
+      }
+    }
+
+    // Return last assistant content
+    const last = [...this.messages].reverse().find(m => m.role === 'assistant')
+    return typeof last?.content === 'string' ? last.content : ''
+  }
+
+  /** Stream a message response. Handles tool calls automatically for multi-turn. */
+  async *stream(message: string): AsyncIterable<StreamChunk> {
+    const config = SainaManager.config
+    const providerName = this._provider ?? config.default
+    const providerConfig = config.providers[providerName]
+    const provider = SainaManager.provider(providerName)
+
+    this.messages.push({ role: 'user', content: message })
+
+    const maxIterations = 10
+    let iterations = 0
+
+    while (iterations < maxIterations) {
+      iterations++
+
+      let fullText = ''
+      const pendingToolCalls: Map<number, { id: string; name: string; args: string }> = new Map()
+
+      for await (const chunk of provider.stream({
+        model: this._model ?? providerConfig?.model ?? '',
+        messages: [...this.messages],
+        system: this._system,
+        tools: this._tools,
+        maxTokens: this._maxTokens ?? config.maxTokens,
+        temperature: this._temperature ?? config.temperature,
+      })) {
+        yield chunk
+
+        if (chunk.type === 'text' && chunk.text) {
+          fullText += chunk.text
+        } else if (chunk.type === 'tool_start' && chunk.toolCall) {
+          pendingToolCalls.set(chunk.toolIndex ?? 0, {
+            id: chunk.toolCall.id ?? '',
+            name: chunk.toolCall.name ?? '',
+            args: '',
+          })
+        } else if (chunk.type === 'tool_delta' && chunk.text) {
+          const pending = pendingToolCalls.get(chunk.toolIndex ?? 0)
+          if (pending) pending.args += chunk.text
+        }
+      }
+
+      // Build tool calls from accumulated stream data
+      const toolCalls: ToolCall[] = []
+      for (const [, pending] of pendingToolCalls) {
+        let args: Record<string, unknown> = {}
+        try {
+          args = JSON.parse(pending.args)
+        } catch {
+          args = pending.args ? { _raw: pending.args } : {}
+        }
+        toolCalls.push({ id: pending.id, name: pending.name, arguments: args })
+      }
+
+      // Append assistant message
+      this.messages.push({
+        role: 'assistant',
+        content: fullText,
+        toolCalls: toolCalls.length > 0 ? toolCalls : undefined,
+      })
+
+      // No tool calls — done
+      if (toolCalls.length === 0) return
+
+      // Execute tools
+      for (const toolCall of toolCalls) {
+        const { message: toolMessage } = await executeTool(this._tools, toolCall)
+        this.messages.push(toolMessage)
+      }
+    }
+  }
+
+  /** Get a copy of all messages in this thread. */
+  getMessages(): Message[] {
+    return [...this.messages]
+  }
+
+  /** Serialize the thread for persistence (session, database, cache). */
+  serialize(): SerializedThread {
+    return {
+      messages: [...this.messages],
+      system: this._system,
+    }
+  }
+
+  /** Restore a previously serialized thread. */
+  restore(data: SerializedThread): this {
+    this.messages = [...data.messages]
+    this._system = data.system
+    return this
+  }
+
+  /** Clear all messages from the thread. */
+  clear(): this {
+    this.messages = []
+    return this
+  }
+}