@garrix82/reactgenie-dsl 1.0.0 → 1.0.2

package/lib/nl/nl-parser.ts

@@ -1,616 +0,0 @@
-import { AgentResponsePayload, PromptGen } from "./prompt-gen";
-import { getLLMMonitor, type LLMMonitor } from "./llm-monitoring";
-import {
-  buildCommandRepairInputText,
-  buildCommandRepairMessages,
-  buildFallbackAgentResponse,
-  CommandValidator,
-  createJsonChatCompletion,
-  DEFAULT_SEMANTIC_PARSER_SAMPLING,
-  extractResponseOutputText,
-  extractStructuredAgentResponse,
-  extractStructuredCommand,
-  getStructuredAgentResponseFormat,
-  getStructuredResponseFormat,
-  getStructuredResponseTextFormat,
-  normalizeAgentResponse,
-  normalizeBaseUrl,
-  SamplingParams,
-} from "./nl-parser-utils";
-export type { SamplingParams } from "./nl-parser-utils";
-
-export interface OpenAIParserOptions {
-  statefulValidationRetry?: boolean;
-}
-
-function buildOpenAIEndpoint(baseUrl: string, suffix: string): string {
-  const normalized = normalizeBaseUrl(baseUrl);
-  if (normalized.endsWith(`/v1/${suffix}`)) {
-    return normalized;
-  }
-  if (normalized.endsWith("/v1")) {
-    return `${normalized}/${suffix}`;
-  }
-  return `${normalized}/v1/${suffix}`;
-}
-
-function requireConfiguredModel(model: string): string {
-  const trimmed = model.trim();
-  if (!trimmed) {
-    throw new Error(
-      "OpenAI parser requires semanticModel / SEMANTIC_MODEL configuration."
-    );
-  }
-  return trimmed;
-}
-
-function asFiniteNumber(value: unknown): number | undefined {
-  return typeof value === "number" && Number.isFinite(value) ? value : undefined;
-}
-
-function extractCostMetrics(response: any): {
-  inputCost?: number;
-  outputCost?: number;
-  totalCost?: number;
-} {
-  const usage = response?.usage ?? response?.usage_metadata ?? {};
-  const inputCost =
-    asFiniteNumber(usage?.input_cost) ??
-    asFiniteNumber(usage?.prompt_cost) ??
-    asFiniteNumber(usage?.prompt_cost_usd);
-  const outputCost =
-    asFiniteNumber(usage?.output_cost) ??
-    asFiniteNumber(usage?.completion_cost) ??
-    asFiniteNumber(usage?.completion_cost_usd);
-  const totalCost =
-    asFiniteNumber(usage?.total_cost) ??
-    asFiniteNumber(usage?.cost) ??
-    asFiniteNumber(usage?.total_cost_usd) ??
-    (typeof inputCost === "number" && typeof outputCost === "number"
-      ? inputCost + outputCost
-      : undefined);
-
-  return {
-    ...(typeof inputCost === "number" ? { inputCost } : {}),
-    ...(typeof outputCost === "number" ? { outputCost } : {}),
-    ...(typeof totalCost === "number" ? { totalCost } : {}),
-  };
-}
-
-export class NlParser {
-  private chatCompletionsEndpoint: string;
-  private responsesEndpoint: string;
-  private samplingParams: Required<SamplingParams>;
-  private llmMonitor!: LLMMonitor;
-  private statefulValidationRetry: boolean;
-
-  constructor(
-    public prompt: PromptGen,
-    private apiKey: string,
-    private basePath: string,
-    private model: string,
-    samplingParams?: SamplingParams,
-    langsmithApiKey?: string,
-    langsmithProject?: string,
-    langsmithEndpoint?: string,
-    private validateCommand?: CommandValidator,
-    parserOptions?: OpenAIParserOptions
-  ) {
-    this.model = requireConfiguredModel(model);
-    this.chatCompletionsEndpoint = buildOpenAIEndpoint(this.basePath, "chat/completions");
-    this.responsesEndpoint = buildOpenAIEndpoint(this.basePath, "responses");
-    this.samplingParams = {
-      ...DEFAULT_SEMANTIC_PARSER_SAMPLING,
-      ...samplingParams,
-    };
-    this.statefulValidationRetry = !!parserOptions?.statefulValidationRetry;
-
-    this.llmMonitor = getLLMMonitor({
-      apiKey: langsmithApiKey,
-      project: langsmithProject,
-      endpoint: langsmithEndpoint,
-    });
-  }
-
-  private async createChatCompletion(body: Record<string, unknown>): Promise<any> {
-    return createJsonChatCompletion({
-      endpoint: this.chatCompletionsEndpoint,
-      apiKey: this.apiKey,
-      body,
-      provider: "openai",
-      requestLabel: "OpenAI-compatible request failed",
-      invalidPayloadLabel: "OpenAI-compatible endpoint returned an invalid JSON payload",
-    });
-  }
-
-  private async createResponse(body: Record<string, unknown>): Promise<any> {
-    return createJsonChatCompletion({
-      endpoint: this.responsesEndpoint,
-      apiKey: this.apiKey,
-      body,
-      provider: "openai",
-      requestLabel: "OpenAI Responses request failed",
-      invalidPayloadLabel: "OpenAI Responses endpoint returned an invalid JSON payload",
-    });
-  }
-
-  private async runStructuredCommand(
-    traceName: string,
-    promptVariants: Array<{ label: string; text: string }>,
-    userUtterance: string
-  ): Promise<{
-    response: any;
-    completion: string;
-    usedPromptText: string;
-    responseId?: string;
-  }> {
-    let usedPromptText = promptVariants[0]?.text || "";
-    let lastError: unknown;
-
-    const runParse = async (
-      strict: boolean,
-      promptText: string
-    ): Promise<{ response: any; command: string; responseId?: string }> => {
-      if (this.statefulValidationRetry) {
-        return this.runStructuredCommandResponse(promptText, strict);
-      }
-      return this.runStructuredCommandMessages([
-        {
-          role: "user",
-          content: promptText,
-        },
-      ], strict);
-    };
-
-    for (const promptVariant of promptVariants) {
-      usedPromptText = promptVariant.text;
-      try {
-        const result = await this.llmMonitor.traceCall(
-          `${traceName}_${promptVariant.label}`,
-          { model: this.model, provider: "openai" },
-          async () => runParse(false, promptVariant.text)
-        );
-        const validated = await this.validateOrRepairCommand(
-          traceName,
-          promptVariant.label,
-          userUtterance,
-          result.command,
-          usedPromptText,
-          result.responseId
-        );
-        return {
-          response: result.response,
-          completion: validated,
-          usedPromptText,
-          ...(result.responseId ? { responseId: result.responseId } : {}),
-        };
-      } catch {
-        try {
-          const result = await this.llmMonitor.traceCall(
-            `${traceName}_${promptVariant.label}`,
-            { model: this.model, provider: "openai" },
-            async () => runParse(true, promptVariant.text)
-          );
-          const validated = await this.validateOrRepairCommand(
-            traceName,
-            promptVariant.label,
-            userUtterance,
-            result.command,
-            usedPromptText,
-            result.responseId
-          );
-          return {
-            response: result.response,
-            completion: validated,
-            usedPromptText,
-            ...(result.responseId ? { responseId: result.responseId } : {}),
-          };
-        } catch (strictError) {
-          lastError = strictError;
-        }
-      }
-    }
-
-    throw (
-      lastError instanceof Error
-        ? lastError
-        : new Error("OpenAI parse failed for filtered prompt")
-    );
-  }
-
-  private async runStructuredCommandResponse(
-    promptText: string,
-    strict: boolean
-  ): Promise<{ response: any; command: string; responseId?: string }> {
-    const response = await this.createResponse({
-      model: this.model,
-      store: true,
-      input: [
-        {
-          role: "user",
-          content: [{ type: "input_text", text: promptText }],
-        },
-      ],
-      text: {
-        format: getStructuredResponseTextFormat(strict),
-      },
-    });
-
-    const raw = extractResponseOutputText(response);
-    const command = extractStructuredCommand(raw);
-    if (!command) {
-      throw new Error("OpenAI Responses structured output missing command");
-    }
-
-    return {
-      response,
-      command,
-      ...(typeof response?.id === "string" ? { responseId: response.id } : {}),
-    };
-  }
-
-  private async runStructuredCommandMessages(
-    messages: ReadonlyArray<{
-      role: "system" | "user" | "assistant";
-      content: string;
-    }>,
-    strict: boolean
-  ): Promise<{ response: any; command: string }> {
-      const response = await this.createChatCompletion({
-        model: this.model,
-        messages: [...messages],
-        temperature: this.samplingParams.temperature,
-        max_tokens: 256,
-        top_p: this.samplingParams.top_p,
-        frequency_penalty: this.samplingParams.frequency_penalty,
-        presence_penalty: this.samplingParams.presence_penalty,
-        response_format: getStructuredResponseFormat(strict),
-      });
-
-      const raw = response.choices[0]?.message?.content ?? null;
-      const command = extractStructuredCommand(raw);
-      if (!command) {
-        throw new Error("OpenAI structured output missing command");
-      }
-      return { response, command };
-  }
-
-  private async validateOrRepairCommand(
-    traceName: string,
-    promptLabel: string,
-    userUtterance: string,
-    command: string,
-    lastUsedPrompt?: string,
-    responseId?: string
-  ): Promise<string> {
-    if (!this.validateCommand) {
-      return command;
-    }
-
-    try {
-      await this.validateCommand(command);
-      return command;
-    } catch (validationError) {
-      const validationMessage =
-        validationError instanceof Error ? validationError.message : String(validationError);
-
-      if (this.statefulValidationRetry && responseId) {
-        let lastError: unknown = validationError;
-        for (const strict of [false, true]) {
-          try {
-            const repaired = await this.llmMonitor.traceCall(
-              `${traceName}_${promptLabel}_repair_${strict ? "strict" : "lenient"}`,
-              { model: this.model, provider: "openai" },
-              async () =>
-                this.runStructuredCommandRepairResponse(
-                  responseId,
-                  buildCommandRepairInputText(userUtterance, command, validationMessage),
-                  strict
-                )
-            );
-            await this.validateCommand(repaired.command);
-            return repaired.command;
-          } catch (repairError) {
-            lastError = repairError;
-          }
-        }
-
-        throw (
-          lastError instanceof Error
-            ? lastError
-            : new Error("OpenAI repair failed after semantic validation")
-        );
-      }
-
-      const messages = buildCommandRepairMessages(
-        userUtterance,
-        command,
-        validationMessage,
-        lastUsedPrompt
-      );
-
-      let lastError: unknown = validationError;
-      for (const strict of [false, true]) {
-        try {
-          const repaired = await this.llmMonitor.traceCall(
-            `${traceName}_${promptLabel}_repair_${strict ? "strict" : "lenient"}`,
-            { model: this.model, provider: "openai" },
-            async () => this.runStructuredCommandMessages(messages, strict)
-          );
-          await this.validateCommand(repaired.command);
-          return repaired.command;
-        } catch (repairError) {
-          lastError = repairError;
-        }
-      }
-
-      throw (
-        lastError instanceof Error
-          ? lastError
-          : new Error("OpenAI repair failed after semantic validation")
-      );
-    }
-  }
-
-  private async runStructuredCommandRepairResponse(
-    previousResponseId: string,
-    repairText: string,
-    strict: boolean
-  ): Promise<{ response: any; command: string; responseId?: string }> {
-    const response = await this.createResponse({
-      model: this.model,
-      store: true,
-      previous_response_id: previousResponseId,
-      input: [
-        {
-          role: "user",
-          content: [{ type: "input_text", text: repairText }],
-        },
-      ],
-      text: {
-        format: getStructuredResponseTextFormat(strict),
-      },
-    });
-
-    const raw = extractResponseOutputText(response);
-    const command = extractStructuredCommand(raw);
-    if (!command) {
-      throw new Error("OpenAI Responses repair output missing command");
-    }
-
-    return {
-      response,
-      command,
-      ...(typeof response?.id === "string" ? { responseId: response.id } : {}),
-    };
-  }
-
-  /** @deprecated Prefer parse() for all new code paths. */
-  async oldParse(nl: string): Promise<string | null> {
-    const promptText = this.prompt.prompt(nl);
-    const startTime = Date.now();
-
-    try {
-      const result = await this.runStructuredCommand("oldParse", [
-        { label: "legacy_prompt", text: promptText },
-      ], nl);
-
-      await this.llmMonitor.logCall({
-        model: this.model,
-        provider: "openai",
-        promptTokens: result.response.usage?.prompt_tokens,
-        completionTokens: result.response.usage?.completion_tokens,
-        totalTokens: result.response.usage?.total_tokens,
-        ...extractCostMetrics(result.response),
-        prompt: result.usedPromptText,
-        completion: result.completion,
-        latency: Date.now() - startTime,
-        timestamp: new Date(),
-      });
-
-      return result.completion;
-    } catch (error) {
-      await this.llmMonitor.logCall({
-        model: this.model,
-        provider: "openai",
-        prompt: promptText,
-        completion: null,
-        latency: Date.now() - startTime,
-        timestamp: new Date(),
-        error: error instanceof Error ? error.message : String(error),
-      });
-      throw error;
-    }
-  }
-
-  async parse(nl: string): Promise<string | null> {
-    const startTime = Date.now();
-    const promptVariants: Array<{ label: string; text: string }> = [];
-    const defaultPrompt = this.prompt.zero_shot_prompt(nl);
-    promptVariants.push({ label: "filtered", text: defaultPrompt });
-
-    try {
-      const result = await this.runStructuredCommand("parse", promptVariants, nl);
-      await this.llmMonitor.logCall({
-        model: this.model,
-        provider: "openai",
-        promptTokens: result.response.usage?.prompt_tokens,
-        completionTokens: result.response.usage?.completion_tokens,
-        totalTokens: result.response.usage?.total_tokens,
-        ...extractCostMetrics(result.response),
-        prompt: result.usedPromptText,
-        completion: result.completion,
-        latency: Date.now() - startTime,
-        timestamp: new Date(),
-      });
-      return result.completion;
-    } catch (error) {
-      await this.llmMonitor.logCall({
-        model: this.model,
-        provider: "openai",
-        prompt: promptVariants[0]?.text || "",
-        completion: null,
-        latency: Date.now() - startTime,
-        timestamp: new Date(),
-        error: error instanceof Error ? error.message : String(error),
-      });
-      throw error;
-    }
-  }
-
-  async parseGpt4(nl: string): Promise<string | null> {
-    const promptText = this.prompt.prompt(nl);
-    const startTime = Date.now();
-
-      const runParse = async (strict: boolean) => {
-        return this.runStructuredCommandMessages(
-          [
-            {
-              role: "system",
-              content: "only generate one line of code",
-            },
-            {
-              role: "user",
-              content: promptText,
-            },
-          ],
-          strict
-        );
-      };
-
-    try {
-      let result;
-      try {
-        result = await this.llmMonitor.traceCall(
-          "parseGpt4",
-          { model: this.model, provider: "openai" },
-          async () => runParse(false)
-        );
-      } catch {
-        result = await this.llmMonitor.traceCall(
-          "parseGpt4",
-          { model: this.model, provider: "openai" },
-          async () => runParse(true)
-        );
-      }
-
-      const validatedCommand = await this.validateOrRepairCommand(
-        "parseGpt4",
-        "legacy_prompt",
-        nl,
-        result.command,
-        promptText
-      );
-
-      await this.llmMonitor.logCall({
-        model: this.model,
-        provider: "openai",
-        promptTokens: result.response.usage?.prompt_tokens,
-        completionTokens: result.response.usage?.completion_tokens,
-        totalTokens: result.response.usage?.total_tokens,
-        ...extractCostMetrics(result.response),
-        prompt: promptText,
-        completion: validatedCommand,
-        latency: Date.now() - startTime,
-        timestamp: new Date(),
-      });
-
-      return validatedCommand;
-    } catch (error) {
-      await this.llmMonitor.logCall({
-        model: this.model,
-        provider: "openai",
-        prompt: promptText,
-        completion: null,
-        latency: Date.now() - startTime,
-        timestamp: new Date(),
-        error: error instanceof Error ? error.message : String(error),
-      });
-      throw error;
-    }
-  }
-
-  async respond(
-    nl: string,
-    parsed: string,
-    result: string
-  ): Promise<AgentResponsePayload | null> {
-    const promptText = this.prompt.response_prompt(nl, parsed, result);
-    const startTime = Date.now();
-
-    const runRespond = async (strict: boolean) => {
-      const response = await this.createChatCompletion({
-        model: this.model,
-        messages: [
-          {
-            role: "system",
-            content: "Return only a strict JSON object for agent_response_json.",
-          },
-          {
-            role: "user",
-            content: promptText,
-          },
-        ],
-        temperature: this.samplingParams.temperature,
-        max_tokens: 256,
-        top_p: this.samplingParams.top_p,
-        frequency_penalty: this.samplingParams.frequency_penalty,
-        presence_penalty: this.samplingParams.presence_penalty,
-        response_format: getStructuredAgentResponseFormat(strict),
-      });
-
-      const raw = response.choices[0]?.message?.content ?? null;
-      const payload = extractStructuredAgentResponse(raw);
-      if (!payload) {
-        throw new Error("OpenAI structured output missing agent response payload");
-      }
-      return { response, payload };
-    };
-
-    try {
-      let result;
-      try {
-        result = await this.llmMonitor.traceCall(
-          "respond",
-          { model: this.model, provider: "openai" },
-          async () => runRespond(false)
-        );
-      } catch {
-        result = await this.llmMonitor.traceCall(
-          "respond",
-          { model: this.model, provider: "openai" },
-          async () => runRespond(true)
-        );
-      }
-
-      const normalizedCompletion = normalizeAgentResponse(result.payload);
-
-      await this.llmMonitor.logCall({
-        model: this.model,
-        provider: "openai",
-        promptTokens: result.response.usage?.prompt_tokens,
-        completionTokens: result.response.usage?.completion_tokens,
-        totalTokens: result.response.usage?.total_tokens,
-        ...extractCostMetrics(result.response),
-        prompt: promptText,
-        completion: JSON.stringify(normalizedCompletion),
-        latency: Date.now() - startTime,
-        timestamp: new Date(),
-      });
-
-      return normalizedCompletion;
-    } catch (error) {
-      const fallbackResponse = buildFallbackAgentResponse(parsed, result);
-      await this.llmMonitor.logCall({
-        model: this.model,
-        provider: "openai",
-        prompt: promptText,
-        completion: JSON.stringify(fallbackResponse),
-        latency: Date.now() - startTime,
-        timestamp: new Date(),
-        error: error instanceof Error ? error.message : String(error),
-      });
-      return fallbackResponse;
-    }
-  }
-}