@mrtdown/core 2.0.0-alpha.2 → 2.0.0-alpha.4

package/dist/llm/functions/extractClaimsFromNewEvidence/index.js

@@ -0,0 +1,153 @@
+import { DateTime } from 'luxon';
+import z from 'zod';
+import { estimateOpenAICostFromUsage, normalizeOpenAIResponsesUsage, } from '#helpers/estimateOpenAICost.js';
+import { openAiClient } from '#llm/client.js';
+import { ClaimSchema } from '#schema/issue/claim.js';
+import { assert } from '#util/assert.js';
+import { buildSystemPrompt } from './prompt.js';
+import { FindLinesTool } from './tools/FindLinesTool.js';
+import { FindServicesTool } from './tools/FindServicesTool.js';
+import { FindStationsTool } from './tools/FindStationsTool.js';
+const TOOL_CALL_LIMIT = 5;
+const ResponseSchema = z.object({
+    claims: z.array(ClaimSchema),
+});
+/**
+ * Extract claims from new evidence.
+ * @param params
+ * @returns
+ */
+export async function extractClaimsFromNewEvidence(params) {
+    const evidenceTs = DateTime.fromISO(params.newEvidence.ts);
+    assert(evidenceTs.isValid, `Invalid date: ${params.newEvidence.ts}`);
+    const findStationsTool = new FindStationsTool(evidenceTs, params.repo);
+    const findServicesTool = new FindServicesTool(evidenceTs, params.repo);
+    const findLinesTool = new FindLinesTool(params.repo);
+    const toolRegistry = {
+        [findStationsTool.name]: findStationsTool,
+        [findServicesTool.name]: findServicesTool,
+        [findLinesTool.name]: findLinesTool,
+    };
+    const context = [
+        {
+            role: 'user',
+            content: `
+Evidence: ${params.newEvidence.text}
+
+Timestamp: ${evidenceTs.toISO({ includeOffset: true })}
+`.trim(),
+        },
+    ];
+    const systemPrompt = buildSystemPrompt();
+    const model = 'gpt-5-mini';
+    let toolCallCount = 0;
+    let response;
+    do {
+        response = await openAiClient.responses.parse({
+            model,
+            instructions: systemPrompt,
+            input: context,
+            reasoning: {
+                effort: 'medium',
+                summary: 'concise',
+            },
+            text: {
+                format: {
+                    type: 'json_schema',
+                    name: 'Response',
+                    strict: true,
+                    schema: z.toJSONSchema(ResponseSchema),
+                },
+            },
+            tools: Object.values(toolRegistry).map((tool) => {
+                return {
+                    type: 'function',
+                    name: tool.name,
+                    description: tool.description,
+                    parameters: tool.paramsSchema,
+                    strict: true,
+                };
+            }),
+            // Don't persist conversation with OpenAI, but include reasoning content to
+            // continue the thread with the same reasoning.
+            store: false,
+            include: ['reasoning.encrypted_content'],
+        });
+        for (const item of response.output) {
+            switch (item.type) {
+                case 'function_call': {
+                    /**
+                     * Prevent the `parsed_arguments` field from being included
+                     * https://github.com/openai/openai-python/issues/2374
+                     */
+                    context.push({
+                        type: 'function_call',
+                        id: item.id,
+                        call_id: item.call_id,
+                        name: item.name,
+                        arguments: item.arguments,
+                    });
+                    if (toolCallCount > TOOL_CALL_LIMIT) {
+                        context.push({
+                            type: 'function_call_output',
+                            call_id: item.call_id,
+                            output: 'Ran out of tool calls. Stop Calling.',
+                        });
+                        console.log('Forced short-circuit, returning error message in tool call result.');
+                    }
+                    if (item.name in toolRegistry) {
+                        const tool = toolRegistry[item.name];
+                        let params;
+                        try {
+                            params = tool.parseParams(JSON.parse(item.arguments));
+                        }
+                        catch (e) {
+                            console.error(`[extractClaimsFromNewEvidence] Error parsing parameters for tool "${item.name}" with arguments "${item.arguments}":`, e);
+                            context.push({
+                                type: 'function_call_output',
+                                call_id: item.call_id,
+                                output: `Invalid parameters for tool "${item.name}". Please try again.`,
+                            });
+                            continue;
+                        }
+                        // Call the tool's run function
+                        const result = await tool.runner(params);
+                        context.push({
+                            type: 'function_call_output',
+                            call_id: item.call_id,
+                            output: result,
+                        });
+                    }
+                    toolCallCount++;
+                    break;
+                }
+                default: {
+                    context.push(item);
+                    break;
+                }
+            }
+        }
+        const usage = normalizeOpenAIResponsesUsage(response.usage);
+        const estimate = estimateOpenAICostFromUsage({ model, usage });
+        if (usage != null) {
+            console.log('[extractClaimsFromNewEvidence] Usage:', {
+                inputTokens: usage.inputTokens,
+                cachedInputTokens: usage.cachedInputTokens,
+                outputTokens: usage.outputTokens,
+                totalTokens: usage.totalTokens,
+            });
+            if (estimate != null) {
+                console.log('[extractClaimsFromNewEvidence] Estimated cost (USD):', estimate.estimatedCostUsd.toFixed(8));
+            }
+            else {
+                console.log(`[extractClaimsFromNewEvidence] No pricing configured for model "${model}".`);
+            }
+        }
+        else {
+            console.log('[extractClaimsFromNewEvidence] Usage is unavailable');
+        }
+    } while (response.output.some((item) => item.type === 'function_call'));
+    assert(response.output_parsed != null, 'Response output parsed is null');
+    return response.output_parsed;
+}
+//# sourceMappingURL=index.js.map

package/dist/llm/functions/extractClaimsFromNewEvidence/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"/","sources":["llm/functions/extractClaimsFromNewEvidence/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,MAAM,OAAO,CAAC;AAKjC,OAAO,CAAC,MAAM,KAAK,CAAC;AACpB,OAAO,EACL,2BAA2B,EAC3B,6BAA6B,GAC9B,MAAM,gCAAgC,CAAC;AACxC,OAAO,EAAE,YAAY,EAAE,MAAM,gBAAgB,CAAC;AAE9C,OAAO,EAAc,WAAW,EAAE,MAAM,wBAAwB,CAAC;AACjE,OAAO,EAAE,MAAM,EAAE,MAAM,iBAAiB,CAAC;AAEzC,OAAO,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAChD,OAAO,EAAE,aAAa,EAAE,MAAM,0BAA0B,CAAC;AACzD,OAAO,EAAE,gBAAgB,EAAE,MAAM,6BAA6B,CAAC;AAC/D,OAAO,EAAE,gBAAgB,EAAE,MAAM,6BAA6B,CAAC;AAE/D,MAAM,eAAe,GAAG,CAAC,CAAC;AAE1B,MAAM,cAAc,GAAG,CAAC,CAAC,MAAM,CAAC;IAC9B,MAAM,EAAE,CAAC,CAAC,KAAK,CAAC,WAAW,CAAC;CAC7B,CAAC,CAAC;AAcH;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,4BAA4B,CAChD,MAA0C;IAE1C,MAAM,UAAU,GAAG,QAAQ,CAAC,OAAO,CAAC,MAAM,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC;IAC3D,MAAM,CAAC,UAAU,CAAC,OAAO,EAAE,iBAAiB,MAAM,CAAC,WAAW,CAAC,EAAE,EAAE,CAAC,CAAC;IAErE,MAAM,gBAAgB,GAAG,IAAI,gBAAgB,CAAC,UAAU,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC;IACvE,MAAM,gBAAgB,GAAG,IAAI,gBAAgB,CAAC,UAAU,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC;IACvE,MAAM,aAAa,GAAG,IAAI,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;IACrD,MAAM,YAAY,GAAiB;QACjC,CAAC,gBAAgB,CAAC,IAAI,CAAC,EAAE,gBAAgB;QACzC,CAAC,gBAAgB,CAAC,IAAI,CAAC,EAAE,gBAAgB;QACzC,CAAC,aAAa,CAAC,IAAI,CAAC,EAAE,aAAa;KACpC,CAAC;IAEF,MAAM,OAAO,GAAwB;QACnC;YACE,IAAI,EAAE,MAAM;YACZ,OAAO,EAAE;YACH,MAAM,CAAC,WAAW,CAAC,IAAI;;aAEtB,UAAU,CAAC,KAAK,CAAC,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC;CACrD,CAAC,IAAI,EAAE;SACH;KACF,CAAC;IAEF,MAAM,YAAY,GAAG,iBAAiB,EAAE,CAAC;IACzC,MAAM,KAAK,GAAG,YAAY,CAAC;IAE3B,IAAI,aAAa,GAAG,CAAC,CAAC;IAEtB,IAAI,QAAwD,CAAC;IAC7D,GAAG,CAAC;QACF,QAAQ,GAAG,MAAM,YAAY,CAAC,SAAS,CAAC,KAAK,CAAC;YAC5C,KAAK;YACL,YAAY,EAAE,YAAY;YAC1B,KAAK,EAAE,OAAO;YACd,SAAS,EAAE;gBACT,MAAM,EAAE,QAAQ;gBAChB,OAAO,EAAE,SAAS;aACnB;YACD,IAAI,EAAE;gBACJ,MAAM,EAAE;oBACN,IAAI,EAAE,aAAa;oBACnB,IAAI,EAAE,UAAU;oBAChB,MAAM,EAAE,IAAI;oBACZ,MAAM,EAAE,CAAC,CAAC,YAAY,CAAC,cAAc,CAAC;iBACvC;aACF;YACD,KAAK,EAAE,MAAM,CAAC,MAAM,CAAC,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE;gBAC9C,OAAO;oBACL,IAAI,EAAE,UAAU;oBAChB,IAAI,EAAE,IAAI,CAAC,IAAI;oBACf,WAAW,EAAE,IAAI,CAAC,WAAW;oBAC7B,UAAU,EAAE,IAAI,CAAC,YAAY;oBAC7B,MAAM,EAAE,IAAI;iBACb,CAAC;YACJ,CAAC,CAAC;YAEF,2EAA2E;YAC3E,+CAA+C;YAC/C,KAAK,EAAE,KAAK;YACZ,OAAO,EAAE,CAAC,6BAA6B,CAAC;SACzC,CAAC,CAAC;QAEH,KAAK,MAAM,IAAI,IAAI,QAAQ,CAAC,MAAM,EAAE,CAAC;YACnC,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;gBAClB,KAAK,eAAe,CAAC,CAAC,CAAC;oBACrB;;;uBAGG;oBACH,OAAO,CAAC,IAAI,CAAC;wBACX,IAAI,EAAE,eAAe;wBACrB,EAAE,EAAE,IAAI,CAAC,EAAE;wBACX,OAAO,EAAE,IAAI,CAAC,OAAO;wBACrB,IAAI,EAAE,IAAI,CAAC,IAAI;wBACf,SAAS,EAAE,IAAI,CAAC,SAAS;qBAC1B,CAAC,CAAC;oBAEH,IAAI,aAAa,GAAG,eAAe,EAAE,CAAC;wBACpC,OAAO,CAAC,IAAI,CAAC;4BACX,IAAI,EAAE,sBAAsB;4BAC5B,OAAO,EAAE,IAAI,CAAC,OAAO;4BACrB,MAAM,EAAE,sCAAsC;yBAC/C,CAAC,CAAC;wBACH,OAAO,CAAC,GAAG,CACT,oEAAoE,CACrE,CAAC;oBACJ,CAAC;oBAED,IAAI,IAAI,CAAC,IAAI,IAAI,YAAY,EAAE,CAAC;wBAC9B,MAAM,IAAI,GAAG,YAAY,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;wBAErC,IAAI,MAAe,CAAC;wBAEpB,IAAI,CAAC;4BACH,MAAM,GAAG,IAAI,CAAC,WAAW,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC;wBACxD,CAAC;wBAAC,OAAO,CAAC,EAAE,CAAC;4BACX,OAAO,CAAC,KAAK,CACX,qEAAqE,IAAI,CAAC,IAAI,qBAAqB,IAAI,CAAC,SAAS,IAAI,EACrH,CAAC,CACF,CAAC;4BACF,OAAO,CAAC,IAAI,CAAC;gCACX,IAAI,EAAE,sBAAsB;gCAC5B,OAAO,EAAE,IAAI,CAAC,OAAO;gCACrB,MAAM,EAAE,gCAAgC,IAAI,CAAC,IAAI,sBAAsB;6BACxE,CAAC,CAAC;4BACH,SAAS;wBACX,CAAC;wBAED,+BAA+B;wBAC/B,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;wBAEzC,OAAO,CAAC,IAAI,CAAC;4BACX,IAAI,EAAE,sBAAsB;4BAC5B,OAAO,EAAE,IAAI,CAAC,OAAO;4BACrB,MAAM,EAAE,MAAM;yBACf,CAAC,CAAC;oBACL,CAAC;oBAED,aAAa,EAAE,CAAC;oBAChB,MAAM;gBACR,CAAC;gBACD,OAAO,CAAC,CAAC,CAAC;oBACR,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;oBACnB,MAAM;gBACR,CAAC;YACH,CAAC;QACH,CAAC;QAED,MAAM,KAAK,GAAG,6BAA6B,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;QAC5D,MAAM,QAAQ,GAAG,2BAA2B,CAAC,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,CAAC;QAC/D,IAAI,KAAK,IAAI,IAAI,EAAE,CAAC;YAClB,OAAO,CAAC,GAAG,CAAC,uCAAuC,EAAE;gBACnD,WAAW,EAAE,KAAK,CAAC,WAAW;gBAC9B,iBAAiB,EAAE,KAAK,CAAC,iBAAiB;gBAC1C,YAAY,EAAE,KAAK,CAAC,YAAY;gBAChC,WAAW,EAAE,KAAK,CAAC,WAAW;aAC/B,CAAC,CAAC;YACH,IAAI,QAAQ,IAAI,IAAI,EAAE,CAAC;gBACrB,OAAO,CAAC,GAAG,CACT,sDAAsD,EACtD,QAAQ,CAAC,gBAAgB,CAAC,OAAO,CAAC,CAAC,CAAC,CACrC,CAAC;YACJ,CAAC;iBAAM,CAAC;gBACN,OAAO,CAAC,GAAG,CACT,mEAAmE,KAAK,IAAI,CAC7E,CAAC;YACJ,CAAC;QACH,CAAC;aAAM,CAAC;YACN,OAAO,CAAC,GAAG,CAAC,qDAAqD,CAAC,CAAC;QACrE,CAAC;IACH,CAAC,QAAQ,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,eAAe,CAAC,EAAE;IAExE,MAAM,CAAC,QAAQ,CAAC,aAAa,IAAI,IAAI,EAAE,gCAAgC,CAAC,CAAC;IAEzE,OAAO,QAAQ,CAAC,aAAa,CAAC;AAChC,CAAC","sourcesContent":["import { DateTime } from 'luxon';\nimport type {\n  ParsedResponse,\n  ResponseInputItem,\n} from 'openai/resources/responses/responses.js';\nimport z from 'zod';\nimport {\n  estimateOpenAICostFromUsage,\n  normalizeOpenAIResponsesUsage,\n} from '#helpers/estimateOpenAICost.js';\nimport { openAiClient } from '#llm/client.js';\nimport type { MRTDownRepository } from '#repo/MRTDownRepository.js';\nimport { type Claim, ClaimSchema } from '#schema/issue/claim.js';\nimport { assert } from '#util/assert.js';\nimport type { ToolRegistry } from '../../common/tool.js';\nimport { buildSystemPrompt } from './prompt.js';\nimport { FindLinesTool } from './tools/FindLinesTool.js';\nimport { FindServicesTool } from './tools/FindServicesTool.js';\nimport { FindStationsTool } from './tools/FindStationsTool.js';\n\nconst TOOL_CALL_LIMIT = 5;\n\nconst ResponseSchema = z.object({\n  claims: z.array(ClaimSchema),\n});\n\nexport interface ExtractClaimsFromNewEvidenceParams {\n  newEvidence: {\n    ts: string;\n    text: string;\n  };\n  repo: MRTDownRepository;\n}\n\nexport type ExtractClaimsFromNewEvidenceResult = {\n  claims: Claim[];\n};\n\n/**\n * Extract claims from new evidence.\n * @param params\n * @returns\n */\nexport async function extractClaimsFromNewEvidence(\n  params: ExtractClaimsFromNewEvidenceParams,\n): Promise<ExtractClaimsFromNewEvidenceResult> {\n  const evidenceTs = DateTime.fromISO(params.newEvidence.ts);\n  assert(evidenceTs.isValid, `Invalid date: ${params.newEvidence.ts}`);\n\n  const findStationsTool = new FindStationsTool(evidenceTs, params.repo);\n  const findServicesTool = new FindServicesTool(evidenceTs, params.repo);\n  const findLinesTool = new FindLinesTool(params.repo);\n  const toolRegistry: ToolRegistry = {\n    [findStationsTool.name]: findStationsTool,\n    [findServicesTool.name]: findServicesTool,\n    [findLinesTool.name]: findLinesTool,\n  };\n\n  const context: ResponseInputItem[] = [\n    {\n      role: 'user',\n      content: `\nEvidence: ${params.newEvidence.text}\n\nTimestamp: ${evidenceTs.toISO({ includeOffset: true })}\n`.trim(),\n    },\n  ];\n\n  const systemPrompt = buildSystemPrompt();\n  const model = 'gpt-5-mini';\n\n  let toolCallCount = 0;\n\n  let response: ParsedResponse<z.infer<typeof ResponseSchema>>;\n  do {\n    response = await openAiClient.responses.parse({\n      model,\n      instructions: systemPrompt,\n      input: context,\n      reasoning: {\n        effort: 'medium',\n        summary: 'concise',\n      },\n      text: {\n        format: {\n          type: 'json_schema',\n          name: 'Response',\n          strict: true,\n          schema: z.toJSONSchema(ResponseSchema),\n        },\n      },\n      tools: Object.values(toolRegistry).map((tool) => {\n        return {\n          type: 'function',\n          name: tool.name,\n          description: tool.description,\n          parameters: tool.paramsSchema,\n          strict: true,\n        };\n      }),\n\n      // Don't persist conversation with OpenAI, but include reasoning content to\n      // continue the thread with the same reasoning.\n      store: false,\n      include: ['reasoning.encrypted_content'],\n    });\n\n    for (const item of response.output) {\n      switch (item.type) {\n        case 'function_call': {\n          /**\n           * Prevent the `parsed_arguments` field from being included\n           * https://github.com/openai/openai-python/issues/2374\n           */\n          context.push({\n            type: 'function_call',\n            id: item.id,\n            call_id: item.call_id,\n            name: item.name,\n            arguments: item.arguments,\n          });\n\n          if (toolCallCount > TOOL_CALL_LIMIT) {\n            context.push({\n              type: 'function_call_output',\n              call_id: item.call_id,\n              output: 'Ran out of tool calls. Stop Calling.',\n            });\n            console.log(\n              'Forced short-circuit, returning error message in tool call result.',\n            );\n          }\n\n          if (item.name in toolRegistry) {\n            const tool = toolRegistry[item.name];\n\n            let params: unknown;\n\n            try {\n              params = tool.parseParams(JSON.parse(item.arguments));\n            } catch (e) {\n              console.error(\n                `[extractClaimsFromNewEvidence] Error parsing parameters for tool \"${item.name}\" with arguments \"${item.arguments}\":`,\n                e,\n              );\n              context.push({\n                type: 'function_call_output',\n                call_id: item.call_id,\n                output: `Invalid parameters for tool \"${item.name}\". Please try again.`,\n              });\n              continue;\n            }\n\n            // Call the tool's run function\n            const result = await tool.runner(params);\n\n            context.push({\n              type: 'function_call_output',\n              call_id: item.call_id,\n              output: result,\n            });\n          }\n\n          toolCallCount++;\n          break;\n        }\n        default: {\n          context.push(item);\n          break;\n        }\n      }\n    }\n\n    const usage = normalizeOpenAIResponsesUsage(response.usage);\n    const estimate = estimateOpenAICostFromUsage({ model, usage });\n    if (usage != null) {\n      console.log('[extractClaimsFromNewEvidence] Usage:', {\n        inputTokens: usage.inputTokens,\n        cachedInputTokens: usage.cachedInputTokens,\n        outputTokens: usage.outputTokens,\n        totalTokens: usage.totalTokens,\n      });\n      if (estimate != null) {\n        console.log(\n          '[extractClaimsFromNewEvidence] Estimated cost (USD):',\n          estimate.estimatedCostUsd.toFixed(8),\n        );\n      } else {\n        console.log(\n          `[extractClaimsFromNewEvidence] No pricing configured for model \"${model}\".`,\n        );\n      }\n    } else {\n      console.log('[extractClaimsFromNewEvidence] Usage is unavailable');\n    }\n  } while (response.output.some((item) => item.type === 'function_call'));\n\n  assert(response.output_parsed != null, 'Response output parsed is null');\n\n  return response.output_parsed;\n}\n"]}

package/dist/llm/functions/extractClaimsFromNewEvidence/prompt.d.ts

@@ -0,0 +1 @@
+export declare function buildSystemPrompt(): string;

package/dist/llm/functions/extractClaimsFromNewEvidence/prompt.js

@@ -0,0 +1,168 @@
+export function buildSystemPrompt() {
+    return `
+You are an expert assistant that extracts structured impact claims from MRT evidence.
+
+Return JSON that matches the provided schema exactly.
+
+## Goal
+Given:
+- Evidence text
+- Evidence timestamp
+
+Extract only operational impact claims (service or station facility impact) that are explicitly stated or strongly implied by the evidence.
+
+## Available tools
+- findLines(lineNames): use to resolve line IDs from line names.
+- findServices(lineId): use to resolve service IDs and service path stations.
+- findStations(stationNames): use to resolve station IDs from station names/codes.
+
+Use tools whenever an ID or station mapping is uncertain. Prefer correctness over guessing.
+
+Important: Line ID is not service ID. A line can have multiple services. For service claims, always use serviceId from findServices; never use lineId from findLines.
+
+Search for valid IDs; never invent or fabricate service IDs, line IDs, or station IDs. When searching for lines or services, try several variations (e.g. "NSL", "NS Line", "North-South Line") as special characters and formatting can be sensitive.
+
+## Claim construction rules
+- One claim per affected entity.
+- Use stable IDs returned by tools.
+- If evidence is non-operational noise (links, generic advisory, no new impact fact), return no claims.
+
+Irrelevance gate (strict):
+- Return claims: [] unless the evidence contains at least one concrete operational assertion about impact state (e.g. delay, no-service, reduced-service, service-hours-adjustment, facility outage/degradation, clear/resume, or explicit planned impact window).
+- Do not generate claims from tags/headers alone (e.g. "[LINE]", "UPDATE", hashtags) without an operational assertion.
+- Advisory-only content (alternative routes, travel advice, support links, "refer to" links, reminders) is irrelevant unless it also includes a concrete impact update.
+- If unsure whether new operational state is stated, prefer no claims.
+
+### entity
+- For train line/service disruptions or maintenance, use:
+  - { type: "service", serviceId }
+  - serviceId must come from findServices (not lineId from findLines). Search for a valid service via findServices; do not make up a service ID.
+- For station facility faults (lift/escalator/screen-door), use:
+  - { type: "facility", stationId, kind }
+
+### statusSignal
+- "open": active disruption/degradation now.
+- "cleared": evidence says issue resolved/resumed/normal service restored.
+- "planned": future scheduled maintenance/planned adjustment.
+
+### effect
+Use object shape with both keys: { service, facility }.
+
+For service entities:
+- Delayed service -> service: { kind: "delay", duration: null unless exact ISO duration is known }
+- Service suspended/closed/no trains -> service: { kind: "no-service" }
+- Reduced frequency/capacity -> service: { kind: "reduced-service" }
+- Temporary operating-hour changes (start later/end earlier/shortened hours) -> service: { kind: "service-hours-adjustment" }
+- If cleared/restored and no ongoing impact remains -> service: null
+- facility must be null
+
+For facility entities:
+- Facility unavailable -> facility: { kind: "facility-out-of-service" }
+- Facility degraded/partially available -> facility: { kind: "facility-degraded" }
+- service must be null
+
+### scopes.service
+- For full line/service statements ("train service resumed", "line closed"), use:
+  - [{ type: "service.whole" }]
+- For "between A and B", use service.segment with station IDs:
+  - { type: "service.segment", fromStationId, toStationId }
+- For station-specific service impact at one station, use:
+  - { type: "service.point", stationId }
+- For service entities, scopes.service should generally be non-null.
+- service.segment is directional: fromStationId -> toStationId is an ordered path, not an unordered pair.
+- Validate segment orientation against the specific service path returned by findServices.
+- For bidirectional output, create one claim per directional service and set segment endpoints to match each service direction (reverse endpoints for reverse direction service).
+- Do not copy the same from/to ordering into both directional services unless that ordering is valid for both service paths.
+
+Direction handling:
+- Priority rule: if direction is ambiguous, default to bidirectional. This rule overrides heuristics.
+- Assume service impact is bidirectional unless direction is stated with strong, unambiguous directional wording.
+- Treat directional impact as valid only when phrasing is truly explicit, for example: "from X to Y towards Z", "towards Z only", "eastbound only", "westbound only", or equivalent unambiguous one-direction wording.
+- "from X to Y" by itself is segment geometry, not directional exclusivity. Do not infer single-direction impact from endpoints alone.
+- Mentions of one segment pair (or one station pair) do not imply one-direction impact by themselves.
+- When direction is not explicitly limited, emit claims for each directional service on the affected line/service.
+- For segment scopes, preserve path direction per service (reverse endpoints for reverse direction service).
+
+### timeHints
+Best effort from evidence. Use null when unknown.
+
+timeHints shape:
+- timeHints is either null or exactly one of:
+  - { kind: "fixed", startAt, endAt }
+  - { kind: "recurring", frequency, startAt, endAt, daysOfWeek, timeWindow, timeZone, excludedDates }
+  - { kind: "start-only", startAt }
+  - { kind: "end-only", endAt }
+- Do not use the old nested shape with startAt/endAt/recurring fields.
+
+Maintenance period rule:
+- For maintenance/planned adjustments, timeHints.startAt/timeHints.endAt must represent the overall maintenance window (calendar period), not the daily active service-impact hours.
+- Put repeated daily/weekly impact hours into recurring.timeWindow.
+- For service-hours-adjustment, recurring.timeWindow must represent when impact applies (restricted/no service), NOT normal operating hours.
+- Example: if temporary operating hours are "first train at 05:45, last train at 23:15", then service runs during 05:45-23:15 and impact interval is 23:15-05:45; set recurring.timeWindow to { startAt: "23:15:00", endAt: "05:45:00" }.
+- Never encode the service-running window as recurring.timeWindow for service-hours-adjustment.
+
+When choosing kind:
+- Use "fixed" when there is a bounded interval (start and maybe end).
+- Use "recurring" when there is a repeating pattern.
+- Use "start-only" when only a start signal is known.
+- Use "end-only" when only an end/clear signal is known.
+
+Field guidance:
+- startAt:
+  - For newly active disruptions, default to evidence timestamp unless a different start time is stated.
+  - For planned items, use the planned start if stated.
+  - For "service will start at X" or "first train at X": the impact window (no service) is from start of day (00:00) until X. Use kind "fixed" with startAt at midnight and endAt at the stated start time.
+- endAt:
+  - For cleared claims with no start in this evidence, prefer kind "end-only" and set endAt to evidence timestamp.
+  - For planned windows, use stated/plausible period end when explicit.
+  - For kind "fixed" and kind "end-only", endAt is exclusive.
+- recurring:
+  - Populate recurring only for repeating patterns.
+  - recurring.startAt and recurring.endAt map to RRULE DTSTART and UNTIL.
+  - RRULE UNTIL semantics: recurring.endAt is the last occurrence anchor (inclusive), not the day after.
+  - Do not default recurring.startAt/endAt to 00:00:00 unless evidence explicitly indicates all-day or midnight boundaries.
+  - Preserve meaningful time-of-day from the described impact window.
+  - The anchor-time rule below is specific to maintenance/infra recurring timeHints:
+    - recurring.startAt/endAt must anchor to the impact START instant (same clock time as recurring.timeWindow.startAt), not recurring.timeWindow.endAt.
+    - For overnight windows, keep start-side anchors across the recurrence range.
+    - Example (non-test): if recurring.timeWindow is { startAt: "22:30:00", endAt: "05:00:00" } on weekdays from 10 Mar to 21 Mar, use startAt/endAt with 22:30 anchors on the first/last applicable dates, not 05:00 anchors.
+  - Shape:
+    - kind: "recurring"
+    - frequency: daily | weekly | monthly | yearly
+    - startAt, endAt as ISO8601 datetimes with offset
+    - daysOfWeek: [MO..SU] or null
+    - timeZone: "Asia/Singapore"
+    - timeWindow: { startAt: "HH:MM:SS", endAt: "HH:MM:SS" }
+    - excludedDates: null unless explicitly stated
+
+## Precision and normalization
+- Prefer explicit facts; avoid inventing causes/details.
+- Use canonical station/service IDs from tools.
+- If only line name is given, map to all relevant services on that line.
+- For ISO8601 datetimes, include timezone offset and seconds. Omit fractional seconds when milliseconds are zero (e.g. use 2026-01-01T07:10:00+08:00, not 2026-01-01T07:10:00.000+08:00).
+- Treat fixed/end-only endAt as exclusive, and recurring.timeWindow.endAt as the end boundary of each daily window.
+- Midnight timestamps are allowed only when explicitly justified by evidence (e.g. "from 00:00", "until midnight", or clear date-only all-day semantics).
+- Keep claims minimal but complete for downstream state updates.
+- Final self-check before returning:
+  - If evidence has no unambiguous one-direction qualifier, ensure claims include all directional services for the affected service/line.
+  - Only return a single directional service claim when explicit direction-only wording is present.
+- Do not include commentary, only schema-conforming JSON.
+
+### causes
+- causes must always be present on every claim.
+- Use null when the evidence does not state or strongly imply a concrete cause subtype.
+- Otherwise set causes to an array containing only valid CauseSubtype enum values.
+- Do not output free-text causes.
+- Conservative mappings:
+  - track/signal/train/power fault -> corresponding *.fault subtype
+  - passenger incident -> passenger.incident
+  - platform screen door fault -> platform_door.fault
+  - engineering/track work -> track.work
+  - upgrade/testing/commissioning -> system.upgrade
+  - lift/escalator outage -> elevator.outage / escalator.outage
+- "maintenance" by itself is not enough to infer track.work; use null unless a concrete subtype is stated.
+- Deterministic rule: if evidence explicitly mentions testing, integrated systems, commissioning, or preparation for a new stage/opening, set causes to include system.upgrade (not null).
+- If multiple independent causes are explicitly stated, include each unique subtype once.
+`.trim();
+}
+//# sourceMappingURL=prompt.js.map

package/dist/llm/functions/extractClaimsFromNewEvidence/prompt.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"prompt.js","sourceRoot":"/","sources":["llm/functions/extractClaimsFromNewEvidence/prompt.ts"],"names":[],"mappings":"AAAA,MAAM,UAAU,iBAAiB;IAC/B,OAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAoKR,CAAC,IAAI,EAAE,CAAC;AACT,CAAC","sourcesContent":["export function buildSystemPrompt() {\n  return `\nYou are an expert assistant that extracts structured impact claims from MRT evidence.\n\nReturn JSON that matches the provided schema exactly.\n\n## Goal\nGiven:\n- Evidence text\n- Evidence timestamp\n\nExtract only operational impact claims (service or station facility impact) that are explicitly stated or strongly implied by the evidence.\n\n## Available tools\n- findLines(lineNames): use to resolve line IDs from line names.\n- findServices(lineId): use to resolve service IDs and service path stations.\n- findStations(stationNames): use to resolve station IDs from station names/codes.\n\nUse tools whenever an ID or station mapping is uncertain. Prefer correctness over guessing.\n\nImportant: Line ID is not service ID. A line can have multiple services. For service claims, always use serviceId from findServices; never use lineId from findLines.\n\nSearch for valid IDs; never invent or fabricate service IDs, line IDs, or station IDs. When searching for lines or services, try several variations (e.g. \"NSL\", \"NS Line\", \"North-South Line\") as special characters and formatting can be sensitive.\n\n## Claim construction rules\n- One claim per affected entity.\n- Use stable IDs returned by tools.\n- If evidence is non-operational noise (links, generic advisory, no new impact fact), return no claims.\n\nIrrelevance gate (strict):\n- Return claims: [] unless the evidence contains at least one concrete operational assertion about impact state (e.g. delay, no-service, reduced-service, service-hours-adjustment, facility outage/degradation, clear/resume, or explicit planned impact window).\n- Do not generate claims from tags/headers alone (e.g. \"[LINE]\", \"UPDATE\", hashtags) without an operational assertion.\n- Advisory-only content (alternative routes, travel advice, support links, \"refer to\" links, reminders) is irrelevant unless it also includes a concrete impact update.\n- If unsure whether new operational state is stated, prefer no claims.\n\n### entity\n- For train line/service disruptions or maintenance, use:\n  - { type: \"service\", serviceId }\n  - serviceId must come from findServices (not lineId from findLines). Search for a valid service via findServices; do not make up a service ID.\n- For station facility faults (lift/escalator/screen-door), use:\n  - { type: \"facility\", stationId, kind }\n\n### statusSignal\n- \"open\": active disruption/degradation now.\n- \"cleared\": evidence says issue resolved/resumed/normal service restored.\n- \"planned\": future scheduled maintenance/planned adjustment.\n\n### effect\nUse object shape with both keys: { service, facility }.\n\nFor service entities:\n- Delayed service -> service: { kind: \"delay\", duration: null unless exact ISO duration is known }\n- Service suspended/closed/no trains -> service: { kind: \"no-service\" }\n- Reduced frequency/capacity -> service: { kind: \"reduced-service\" }\n- Temporary operating-hour changes (start later/end earlier/shortened hours) -> service: { kind: \"service-hours-adjustment\" }\n- If cleared/restored and no ongoing impact remains -> service: null\n- facility must be null\n\nFor facility entities:\n- Facility unavailable -> facility: { kind: \"facility-out-of-service\" }\n- Facility degraded/partially available -> facility: { kind: \"facility-degraded\" }\n- service must be null\n\n### scopes.service\n- For full line/service statements (\"train service resumed\", \"line closed\"), use:\n  - [{ type: \"service.whole\" }]\n- For \"between A and B\", use service.segment with station IDs:\n  - { type: \"service.segment\", fromStationId, toStationId }\n- For station-specific service impact at one station, use:\n  - { type: \"service.point\", stationId }\n- For service entities, scopes.service should generally be non-null.\n- service.segment is directional: fromStationId -> toStationId is an ordered path, not an unordered pair.\n- Validate segment orientation against the specific service path returned by findServices.\n- For bidirectional output, create one claim per directional service and set segment endpoints to match each service direction (reverse endpoints for reverse direction service).\n- Do not copy the same from/to ordering into both directional services unless that ordering is valid for both service paths.\n\nDirection handling:\n- Priority rule: if direction is ambiguous, default to bidirectional. This rule overrides heuristics.\n- Assume service impact is bidirectional unless direction is stated with strong, unambiguous directional wording.\n- Treat directional impact as valid only when phrasing is truly explicit, for example: \"from X to Y towards Z\", \"towards Z only\", \"eastbound only\", \"westbound only\", or equivalent unambiguous one-direction wording.\n- \"from X to Y\" by itself is segment geometry, not directional exclusivity. Do not infer single-direction impact from endpoints alone.\n- Mentions of one segment pair (or one station pair) do not imply one-direction impact by themselves.\n- When direction is not explicitly limited, emit claims for each directional service on the affected line/service.\n- For segment scopes, preserve path direction per service (reverse endpoints for reverse direction service).\n\n### timeHints\nBest effort from evidence. Use null when unknown.\n\ntimeHints shape:\n- timeHints is either null or exactly one of:\n  - { kind: \"fixed\", startAt, endAt }\n  - { kind: \"recurring\", frequency, startAt, endAt, daysOfWeek, timeWindow, timeZone, excludedDates }\n  - { kind: \"start-only\", startAt }\n  - { kind: \"end-only\", endAt }\n- Do not use the old nested shape with startAt/endAt/recurring fields.\n\nMaintenance period rule:\n- For maintenance/planned adjustments, timeHints.startAt/timeHints.endAt must represent the overall maintenance window (calendar period), not the daily active service-impact hours.\n- Put repeated daily/weekly impact hours into recurring.timeWindow.\n- For service-hours-adjustment, recurring.timeWindow must represent when impact applies (restricted/no service), NOT normal operating hours.\n- Example: if temporary operating hours are \"first train at 05:45, last train at 23:15\", then service runs during 05:45-23:15 and impact interval is 23:15-05:45; set recurring.timeWindow to { startAt: \"23:15:00\", endAt: \"05:45:00\" }.\n- Never encode the service-running window as recurring.timeWindow for service-hours-adjustment.\n\nWhen choosing kind:\n- Use \"fixed\" when there is a bounded interval (start and maybe end).\n- Use \"recurring\" when there is a repeating pattern.\n- Use \"start-only\" when only a start signal is known.\n- Use \"end-only\" when only an end/clear signal is known.\n\nField guidance:\n- startAt:\n  - For newly active disruptions, default to evidence timestamp unless a different start time is stated.\n  - For planned items, use the planned start if stated.\n  - For \"service will start at X\" or \"first train at X\": the impact window (no service) is from start of day (00:00) until X. Use kind \"fixed\" with startAt at midnight and endAt at the stated start time.\n- endAt:\n  - For cleared claims with no start in this evidence, prefer kind \"end-only\" and set endAt to evidence timestamp.\n  - For planned windows, use stated/plausible period end when explicit.\n  - For kind \"fixed\" and kind \"end-only\", endAt is exclusive.\n- recurring:\n  - Populate recurring only for repeating patterns.\n  - recurring.startAt and recurring.endAt map to RRULE DTSTART and UNTIL.\n  - RRULE UNTIL semantics: recurring.endAt is the last occurrence anchor (inclusive), not the day after.\n  - Do not default recurring.startAt/endAt to 00:00:00 unless evidence explicitly indicates all-day or midnight boundaries.\n  - Preserve meaningful time-of-day from the described impact window.\n  - The anchor-time rule below is specific to maintenance/infra recurring timeHints:\n    - recurring.startAt/endAt must anchor to the impact START instant (same clock time as recurring.timeWindow.startAt), not recurring.timeWindow.endAt.\n    - For overnight windows, keep start-side anchors across the recurrence range.\n    - Example (non-test): if recurring.timeWindow is { startAt: \"22:30:00\", endAt: \"05:00:00\" } on weekdays from 10 Mar to 21 Mar, use startAt/endAt with 22:30 anchors on the first/last applicable dates, not 05:00 anchors.\n  - Shape:\n    - kind: \"recurring\"\n    - frequency: daily | weekly | monthly | yearly\n    - startAt, endAt as ISO8601 datetimes with offset\n    - daysOfWeek: [MO..SU] or null\n    - timeZone: \"Asia/Singapore\"\n    - timeWindow: { startAt: \"HH:MM:SS\", endAt: \"HH:MM:SS\" }\n    - excludedDates: null unless explicitly stated\n\n## Precision and normalization\n- Prefer explicit facts; avoid inventing causes/details.\n- Use canonical station/service IDs from tools.\n- If only line name is given, map to all relevant services on that line.\n- For ISO8601 datetimes, include timezone offset and seconds. Omit fractional seconds when milliseconds are zero (e.g. use 2026-01-01T07:10:00+08:00, not 2026-01-01T07:10:00.000+08:00).\n- Treat fixed/end-only endAt as exclusive, and recurring.timeWindow.endAt as the end boundary of each daily window.\n- Midnight timestamps are allowed only when explicitly justified by evidence (e.g. \"from 00:00\", \"until midnight\", or clear date-only all-day semantics).\n- Keep claims minimal but complete for downstream state updates.\n- Final self-check before returning:\n  - If evidence has no unambiguous one-direction qualifier, ensure claims include all directional services for the affected service/line.\n  - Only return a single directional service claim when explicit direction-only wording is present.\n- Do not include commentary, only schema-conforming JSON.\n\n### causes\n- causes must always be present on every claim.\n- Use null when the evidence does not state or strongly imply a concrete cause subtype.\n- Otherwise set causes to an array containing only valid CauseSubtype enum values.\n- Do not output free-text causes.\n- Conservative mappings:\n  - track/signal/train/power fault -> corresponding *.fault subtype\n  - passenger incident -> passenger.incident\n  - platform screen door fault -> platform_door.fault\n  - engineering/track work -> track.work\n  - upgrade/testing/commissioning -> system.upgrade\n  - lift/escalator outage -> elevator.outage / escalator.outage\n- \"maintenance\" by itself is not enough to infer track.work; use null unless a concrete subtype is stated.\n- Deterministic rule: if evidence explicitly mentions testing, integrated systems, commissioning, or preparation for a new stage/opening, set causes to include system.upgrade (not null).\n- If multiple independent causes are explicitly stated, include each unique subtype once.\n`.trim();\n}\n"]}

package/dist/llm/functions/extractClaimsFromNewEvidence/tools/FindLinesTool.d.ts

@@ -0,0 +1,19 @@
+import z from 'zod';
+import type { MRTDownRepository } from '#repo/MRTDownRepository.js';
+import { Tool } from '../../../common/tool.js';
+declare const FindLinesToolParametersSchema: z.ZodObject<{
+    lineNames: z.ZodArray<z.ZodString>;
+}, z.z.core.$strip>;
+type FindLinesToolParameters = z.infer<typeof FindLinesToolParametersSchema>;
+export declare class FindLinesTool extends Tool<FindLinesToolParameters> {
+    name: string;
+    description: string;
+    private readonly repo;
+    constructor(repo: MRTDownRepository);
+    get paramsSchema(): {
+        [key: string]: unknown;
+    };
+    parseParams(params: unknown): FindLinesToolParameters;
+    runner(params: FindLinesToolParameters): Promise<string>;
+}
+export {};

package/dist/llm/functions/extractClaimsFromNewEvidence/tools/FindLinesTool.js

@@ -0,0 +1,65 @@
+import { gfmToMarkdown } from 'mdast-util-gfm';
+import { toMarkdown } from 'mdast-util-to-markdown';
+import z from 'zod';
+import { Tool } from '../../../common/tool.js';
+const FindLinesToolParametersSchema = z.object({
+    lineNames: z.array(z.string()),
+});
+export class FindLinesTool extends Tool {
+    name = 'findLines';
+    description = 'Find lines by name';
+    repo;
+    constructor(repo) {
+        super();
+        this.repo = repo;
+    }
+    get paramsSchema() {
+        return z.toJSONSchema(FindLinesToolParametersSchema);
+    }
+    parseParams(params) {
+        return FindLinesToolParametersSchema.parse(params);
+    }
+    async runner(params) {
+        console.log('[findLines] Calling tool with parameters:', params);
+        const lines = this.repo.lines.searchByName(params.lineNames);
+        const table = {
+            type: 'table',
+            children: [
+                {
+                    type: 'tableRow',
+                    children: [
+                        {
+                            type: 'tableCell',
+                            children: [{ type: 'text', value: 'Line ID' }],
+                        },
+                        {
+                            type: 'tableCell',
+                            children: [{ type: 'text', value: 'Line Name' }],
+                        },
+                    ],
+                },
+            ],
+        };
+        for (const line of lines) {
+            table.children.push({
+                type: 'tableRow',
+                children: [
+                    {
+                        type: 'tableCell',
+                        children: [{ type: 'text', value: line.id }],
+                    },
+                    {
+                        type: 'tableCell',
+                        children: [{ type: 'text', value: line.name['en-SG'] }],
+                    },
+                ],
+            });
+        }
+        const output = toMarkdown(table, {
+            extensions: [gfmToMarkdown()],
+        });
+        console.log(`[findLines] Response output:\n${output}`);
+        return output;
+    }
+}
+//# sourceMappingURL=FindLinesTool.js.map

package/dist/llm/functions/extractClaimsFromNewEvidence/tools/FindLinesTool.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"FindLinesTool.js","sourceRoot":"/","sources":["llm/functions/extractClaimsFromNewEvidence/tools/FindLinesTool.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAC/C,OAAO,EAAE,UAAU,EAAE,MAAM,wBAAwB,CAAC;AACpD,OAAO,CAAC,MAAM,KAAK,CAAC;AAEpB,OAAO,EAAE,IAAI,EAAE,MAAM,yBAAyB,CAAC;AAE/C,MAAM,6BAA6B,GAAG,CAAC,CAAC,MAAM,CAAC;IAC7C,SAAS,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC;CAC/B,CAAC,CAAC;AAGH,MAAM,OAAO,aAAc,SAAQ,IAA6B;IACvD,IAAI,GAAG,WAAW,CAAC;IACnB,WAAW,GAAG,oBAAoB,CAAC;IACzB,IAAI,CAAoB;IAEzC,YAAY,IAAuB;QACjC,KAAK,EAAE,CAAC;QACR,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;IACnB,CAAC;IAED,IAAW,YAAY;QACrB,OAAO,CAAC,CAAC,YAAY,CAAC,6BAA6B,CAAC,CAAC;IACvD,CAAC;IAEM,WAAW,CAAC,MAAe;QAChC,OAAO,6BAA6B,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;IACrD,CAAC;IAEM,KAAK,CAAC,MAAM,CAAC,MAA+B;QACjD,OAAO,CAAC,GAAG,CAAC,2CAA2C,EAAE,MAAM,CAAC,CAAC;QAEjE,MAAM,KAAK,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC;QAE7D,MAAM,KAAK,GAAU;YACnB,IAAI,EAAE,OAAO;YACb,QAAQ,EAAE;gBACR;oBACE,IAAI,EAAE,UAAU;oBAChB,QAAQ,EAAE;wBACR;4BACE,IAAI,EAAE,WAAW;4BACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,SAAS,EAAE,CAAC;yBAC/C;wBACD;4BACE,IAAI,EAAE,WAAW;4BACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,WAAW,EAAE,CAAC;yBACjD;qBACF;iBACF;aACF;SACF,CAAC;QAEF,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;YACzB,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC;gBAClB,IAAI,EAAE,UAAU;gBAChB,QAAQ,EAAE;oBACR;wBACE,IAAI,EAAE,WAAW;wBACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC;qBAC7C;oBACD;wBACE,IAAI,EAAE,WAAW;wBACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;qBACxD;iBACF;aACF,CAAC,CAAC;QACL,CAAC;QAED,MAAM,MAAM,GAAG,UAAU,CAAC,KAAK,EAAE;YAC/B,UAAU,EAAE,CAAC,aAAa,EAAE,CAAC;SAC9B,CAAC,CAAC;QACH,OAAO,CAAC,GAAG,CAAC,iCAAiC,MAAM,EAAE,CAAC,CAAC;QAEvD,OAAO,MAAM,CAAC;IAChB,CAAC;CACF","sourcesContent":["import type { Table } from 'mdast';\nimport { gfmToMarkdown } from 'mdast-util-gfm';\nimport { toMarkdown } from 'mdast-util-to-markdown';\nimport z from 'zod';\nimport type { MRTDownRepository } from '#repo/MRTDownRepository.js';\nimport { Tool } from '../../../common/tool.js';\n\nconst FindLinesToolParametersSchema = z.object({\n  lineNames: z.array(z.string()),\n});\ntype FindLinesToolParameters = z.infer<typeof FindLinesToolParametersSchema>;\n\nexport class FindLinesTool extends Tool<FindLinesToolParameters> {\n  public name = 'findLines';\n  public description = 'Find lines by name';\n  private readonly repo: MRTDownRepository;\n\n  constructor(repo: MRTDownRepository) {\n    super();\n    this.repo = repo;\n  }\n\n  public get paramsSchema(): { [key: string]: unknown } {\n    return z.toJSONSchema(FindLinesToolParametersSchema);\n  }\n\n  public parseParams(params: unknown): FindLinesToolParameters {\n    return FindLinesToolParametersSchema.parse(params);\n  }\n\n  public async runner(params: FindLinesToolParameters): Promise<string> {\n    console.log('[findLines] Calling tool with parameters:', params);\n\n    const lines = this.repo.lines.searchByName(params.lineNames);\n\n    const table: Table = {\n      type: 'table',\n      children: [\n        {\n          type: 'tableRow',\n          children: [\n            {\n              type: 'tableCell',\n              children: [{ type: 'text', value: 'Line ID' }],\n            },\n            {\n              type: 'tableCell',\n              children: [{ type: 'text', value: 'Line Name' }],\n            },\n          ],\n        },\n      ],\n    };\n\n    for (const line of lines) {\n      table.children.push({\n        type: 'tableRow',\n        children: [\n          {\n            type: 'tableCell',\n            children: [{ type: 'text', value: line.id }],\n          },\n          {\n            type: 'tableCell',\n            children: [{ type: 'text', value: line.name['en-SG'] }],\n          },\n        ],\n      });\n    }\n\n    const output = toMarkdown(table, {\n      extensions: [gfmToMarkdown()],\n    });\n    console.log(`[findLines] Response output:\\n${output}`);\n\n    return output;\n  }\n}\n"]}

package/dist/llm/functions/extractClaimsFromNewEvidence/tools/FindServicesTool.d.ts

@@ -0,0 +1,21 @@
+import { DateTime } from 'luxon';
+import z from 'zod';
+import type { MRTDownRepository } from '#repo/MRTDownRepository.js';
+import { Tool } from '../../../common/tool.js';
+declare const FindServicesToolParametersSchema: z.ZodObject<{
+    lineId: z.ZodString;
+}, z.z.core.$strip>;
+type FindServicesToolParameters = z.infer<typeof FindServicesToolParametersSchema>;
+export declare class FindServicesTool extends Tool<FindServicesToolParameters> {
+    name: string;
+    description: string;
+    private readonly evidenceTs;
+    private readonly repo;
+    constructor(evidenceTs: DateTime, repo: MRTDownRepository);
+    get paramsSchema(): {
+        [key: string]: unknown;
+    };
+    parseParams(params: unknown): FindServicesToolParameters;
+    runner(params: FindServicesToolParameters): Promise<string>;
+}
+export {};

package/dist/llm/functions/extractClaimsFromNewEvidence/tools/FindServicesTool.js

@@ -0,0 +1,115 @@
+import { DateTime } from 'luxon';
+import { gfmToMarkdown } from 'mdast-util-gfm';
+import { toMarkdown } from 'mdast-util-to-markdown';
+import z from 'zod';
+import { assert } from '#util/assert.js';
+import { Tool } from '../../../common/tool.js';
+const FindServicesToolParametersSchema = z.object({
+    lineId: z.string(),
+});
+export class FindServicesTool extends Tool {
+    name = 'findServices';
+    description = 'Find services by name';
+    evidenceTs;
+    repo;
+    constructor(evidenceTs, repo) {
+        super();
+        this.evidenceTs = evidenceTs;
+        this.repo = repo;
+    }
+    get paramsSchema() {
+        return z.toJSONSchema(FindServicesToolParametersSchema);
+    }
+    parseParams(params) {
+        return FindServicesToolParametersSchema.parse(params);
+    }
+    async runner(params) {
+        console.log('[findServices] Calling tool with parameters:', params);
+        const services = this.repo.services.searchByLineId(params.lineId);
+        const table = {
+            type: 'table',
+            children: [
+                {
+                    type: 'tableRow',
+                    children: [
+                        {
+                            type: 'tableCell',
+                            children: [{ type: 'text', value: 'Service ID' }],
+                        },
+                        {
+                            type: 'tableCell',
+                            children: [{ type: 'text', value: 'Service Name' }],
+                        },
+                        {
+                            type: 'tableCell',
+                            children: [{ type: 'text', value: 'Line ID' }],
+                        },
+                        {
+                            type: 'tableCell',
+                            children: [{ type: 'text', value: 'Station IDs' }],
+                        },
+                        {
+                            type: 'tableCell',
+                            children: [{ type: 'text', value: 'Operating Hours' }],
+                        },
+                    ],
+                },
+            ],
+        };
+        for (const service of services) {
+            const relevantRevision = service.revisions.findLast((revision) => {
+                const startAt = DateTime.fromISO(revision.startAt);
+                assert(startAt.isValid, `Invalid date: ${revision.startAt}`);
+                if (revision.endAt == null) {
+                    return startAt <= this.evidenceTs;
+                }
+                const endAt = DateTime.fromISO(revision.endAt);
+                assert(endAt.isValid, `Invalid date: ${revision.endAt}`);
+                return startAt <= this.evidenceTs && endAt > this.evidenceTs;
+            });
+            if (relevantRevision == null)
+                continue;
+            table.children.push({
+                type: 'tableRow',
+                children: [
+                    {
+                        type: 'tableCell',
+                        children: [{ type: 'text', value: service.id }],
+                    },
+                    {
+                        type: 'tableCell',
+                        children: [{ type: 'text', value: service.name['en-SG'] }],
+                    },
+                    {
+                        type: 'tableCell',
+                        children: [{ type: 'text', value: service.lineId }],
+                    },
+                    {
+                        type: 'tableCell',
+                        children: [
+                            {
+                                type: 'text',
+                                value: `${relevantRevision.path.stations.length} (${relevantRevision.path.stations.map((station) => station.stationId).join('→')})`,
+                            },
+                        ],
+                    },
+                    {
+                        type: 'tableCell',
+                        children: [
+                            {
+                                type: 'text',
+                                value: `Weekdays: ${relevantRevision.operatingHours.weekdays.start}-${relevantRevision.operatingHours.weekdays.end} | Weekends: ${relevantRevision.operatingHours.weekends.start}-${relevantRevision.operatingHours.weekends.end}`,
+                            },
+                        ],
+                    },
+                ],
+            });
+        }
+        const output = toMarkdown(table, {
+            extensions: [gfmToMarkdown()],
+        });
+        console.log(`[findServices] Response output:\n${output}`);
+        return output;
+    }
+}
+//# sourceMappingURL=FindServicesTool.js.map

package/dist/llm/functions/extractClaimsFromNewEvidence/tools/FindServicesTool.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"FindServicesTool.js","sourceRoot":"/","sources":["llm/functions/extractClaimsFromNewEvidence/tools/FindServicesTool.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,MAAM,OAAO,CAAC;AAEjC,OAAO,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAC/C,OAAO,EAAE,UAAU,EAAE,MAAM,wBAAwB,CAAC;AACpD,OAAO,CAAC,MAAM,KAAK,CAAC;AAEpB,OAAO,EAAE,MAAM,EAAE,MAAM,iBAAiB,CAAC;AACzC,OAAO,EAAE,IAAI,EAAE,MAAM,yBAAyB,CAAC;AAE/C,MAAM,gCAAgC,GAAG,CAAC,CAAC,MAAM,CAAC;IAChD,MAAM,EAAE,CAAC,CAAC,MAAM,EAAE;CACnB,CAAC,CAAC;AAKH,MAAM,OAAO,gBAAiB,SAAQ,IAAgC;IAC7D,IAAI,GAAG,cAAc,CAAC;IACtB,WAAW,GAAG,uBAAuB,CAAC;IAC5B,UAAU,CAAW;IACrB,IAAI,CAAoB;IAEzC,YAAY,UAAoB,EAAE,IAAuB;QACvD,KAAK,EAAE,CAAC;QACR,IAAI,CAAC,UAAU,GAAG,UAAU,CAAC;QAC7B,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;IACnB,CAAC;IAED,IAAW,YAAY;QACrB,OAAO,CAAC,CAAC,YAAY,CAAC,gCAAgC,CAAC,CAAC;IAC1D,CAAC;IAEM,WAAW,CAAC,MAAe;QAChC,OAAO,gCAAgC,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;IACxD,CAAC;IAEM,KAAK,CAAC,MAAM,CAAC,MAAkC;QACpD,OAAO,CAAC,GAAG,CAAC,8CAA8C,EAAE,MAAM,CAAC,CAAC;QAEpE,MAAM,QAAQ,GAAG,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,cAAc,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;QAElE,MAAM,KAAK,GAAU;YACnB,IAAI,EAAE,OAAO;YACb,QAAQ,EAAE;gBACR;oBACE,IAAI,EAAE,UAAU;oBAChB,QAAQ,EAAE;wBACR;4BACE,IAAI,EAAE,WAAW;4BACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,YAAY,EAAE,CAAC;yBAClD;wBACD;4BACE,IAAI,EAAE,WAAW;4BACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,cAAc,EAAE,CAAC;yBACpD;wBACD;4BACE,IAAI,EAAE,WAAW;4BACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,SAAS,EAAE,CAAC;yBAC/C;wBACD;4BACE,IAAI,EAAE,WAAW;4BACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,aAAa,EAAE,CAAC;yBACnD;wBACD;4BACE,IAAI,EAAE,WAAW;4BACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,iBAAiB,EAAE,CAAC;yBACvD;qBACF;iBACF;aACF;SACF,CAAC;QAEF,KAAK,MAAM,OAAO,IAAI,QAAQ,EAAE,CAAC;YAC/B,MAAM,gBAAgB,GAAG,OAAO,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,QAAQ,EAAE,EAAE;gBAC/D,MAAM,OAAO,GAAG,QAAQ,CAAC,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;gBACnD,MAAM,CAAC,OAAO,CAAC,OAAO,EAAE,iBAAiB,QAAQ,CAAC,OAAO,EAAE,CAAC,CAAC;gBAE7D,IAAI,QAAQ,CAAC,KAAK,IAAI,IAAI,EAAE,CAAC;oBAC3B,OAAO,OAAO,IAAI,IAAI,CAAC,UAAU,CAAC;gBACpC,CAAC;gBAED,MAAM,KAAK,GAAG,QAAQ,CAAC,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;gBAC/C,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,iBAAiB,QAAQ,CAAC,KAAK,EAAE,CAAC,CAAC;gBAEzD,OAAO,OAAO,IAAI,IAAI,CAAC,UAAU,IAAI,KAAK,GAAG,IAAI,CAAC,UAAU,CAAC;YAC/D,CAAC,CAAC,CAAC;YAEH,IAAI,gBAAgB,IAAI,IAAI;gBAAE,SAAS;YAEvC,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC;gBAClB,IAAI,EAAE,UAAU;gBAChB,QAAQ,EAAE;oBACR;wBACE,IAAI,EAAE,WAAW;wBACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,OAAO,CAAC,EAAE,EAAE,CAAC;qBAChD;oBACD;wBACE,IAAI,EAAE,WAAW;wBACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;qBAC3D;oBACD;wBACE,IAAI,EAAE,WAAW;wBACjB,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC;qBACpD;oBACD;wBACE,IAAI,EAAE,WAAW;wBACjB,QAAQ,EAAE;4BACR;gCACE,IAAI,EAAE,MAAM;gCACZ,KAAK,EAAE,GAAG,gBAAgB,CAAC,IAAI,CAAC,QAAQ,CAAC,MAAM,KAAK,gBAAgB,CAAC,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,GAAG;6BACpI;yBACF;qBACF;oBACD;wBACE,IAAI,EAAE,WAAW;wBACjB,QAAQ,EAAE;4BACR;gCACE,IAAI,EAAE,MAAM;gCACZ,KAAK,EAAE,aAAa,gBAAgB,CAAC,cAAc,CAAC,QAAQ,CAAC,KAAK,IAAI,gBAAgB,CAAC,cAAc,CAAC,QAAQ,CAAC,GAAG,gBAAgB,gBAAgB,CAAC,cAAc,CAAC,QAAQ,CAAC,KAAK,IAAI,gBAAgB,CAAC,cAAc,CAAC,QAAQ,CAAC,GAAG,EAAE;6BACnO;yBACF;qBACF;iBACF;aACF,CAAC,CAAC;QACL,CAAC;QAED,MAAM,MAAM,GAAG,UAAU,CAAC,KAAK,EAAE;YAC/B,UAAU,EAAE,CAAC,aAAa,EAAE,CAAC;SAC9B,CAAC,CAAC;QACH,OAAO,CAAC,GAAG,CAAC,oCAAoC,MAAM,EAAE,CAAC,CAAC;QAE1D,OAAO,MAAM,CAAC;IAChB,CAAC;CACF","sourcesContent":["import { DateTime } from 'luxon';\nimport type { Table } from 'mdast';\nimport { gfmToMarkdown } from 'mdast-util-gfm';\nimport { toMarkdown } from 'mdast-util-to-markdown';\nimport z from 'zod';\nimport type { MRTDownRepository } from '#repo/MRTDownRepository.js';\nimport { assert } from '#util/assert.js';\nimport { Tool } from '../../../common/tool.js';\n\nconst FindServicesToolParametersSchema = z.object({\n  lineId: z.string(),\n});\ntype FindServicesToolParameters = z.infer<\n  typeof FindServicesToolParametersSchema\n>;\n\nexport class FindServicesTool extends Tool<FindServicesToolParameters> {\n  public name = 'findServices';\n  public description = 'Find services by name';\n  private readonly evidenceTs: DateTime;\n  private readonly repo: MRTDownRepository;\n\n  constructor(evidenceTs: DateTime, repo: MRTDownRepository) {\n    super();\n    this.evidenceTs = evidenceTs;\n    this.repo = repo;\n  }\n\n  public get paramsSchema(): { [key: string]: unknown } {\n    return z.toJSONSchema(FindServicesToolParametersSchema);\n  }\n\n  public parseParams(params: unknown): FindServicesToolParameters {\n    return FindServicesToolParametersSchema.parse(params);\n  }\n\n  public async runner(params: FindServicesToolParameters): Promise<string> {\n    console.log('[findServices] Calling tool with parameters:', params);\n\n    const services = this.repo.services.searchByLineId(params.lineId);\n\n    const table: Table = {\n      type: 'table',\n      children: [\n        {\n          type: 'tableRow',\n          children: [\n            {\n              type: 'tableCell',\n              children: [{ type: 'text', value: 'Service ID' }],\n            },\n            {\n              type: 'tableCell',\n              children: [{ type: 'text', value: 'Service Name' }],\n            },\n            {\n              type: 'tableCell',\n              children: [{ type: 'text', value: 'Line ID' }],\n            },\n            {\n              type: 'tableCell',\n              children: [{ type: 'text', value: 'Station IDs' }],\n            },\n            {\n              type: 'tableCell',\n              children: [{ type: 'text', value: 'Operating Hours' }],\n            },\n          ],\n        },\n      ],\n    };\n\n    for (const service of services) {\n      const relevantRevision = service.revisions.findLast((revision) => {\n        const startAt = DateTime.fromISO(revision.startAt);\n        assert(startAt.isValid, `Invalid date: ${revision.startAt}`);\n\n        if (revision.endAt == null) {\n          return startAt <= this.evidenceTs;\n        }\n\n        const endAt = DateTime.fromISO(revision.endAt);\n        assert(endAt.isValid, `Invalid date: ${revision.endAt}`);\n\n        return startAt <= this.evidenceTs && endAt > this.evidenceTs;\n      });\n\n      if (relevantRevision == null) continue;\n\n      table.children.push({\n        type: 'tableRow',\n        children: [\n          {\n            type: 'tableCell',\n            children: [{ type: 'text', value: service.id }],\n          },\n          {\n            type: 'tableCell',\n            children: [{ type: 'text', value: service.name['en-SG'] }],\n          },\n          {\n            type: 'tableCell',\n            children: [{ type: 'text', value: service.lineId }],\n          },\n          {\n            type: 'tableCell',\n            children: [\n              {\n                type: 'text',\n                value: `${relevantRevision.path.stations.length} (${relevantRevision.path.stations.map((station) => station.stationId).join('→')})`,\n              },\n            ],\n          },\n          {\n            type: 'tableCell',\n            children: [\n              {\n                type: 'text',\n                value: `Weekdays: ${relevantRevision.operatingHours.weekdays.start}-${relevantRevision.operatingHours.weekdays.end} | Weekends: ${relevantRevision.operatingHours.weekends.start}-${relevantRevision.operatingHours.weekends.end}`,\n              },\n            ],\n          },\n        ],\n      });\n    }\n\n    const output = toMarkdown(table, {\n      extensions: [gfmToMarkdown()],\n    });\n    console.log(`[findServices] Response output:\\n${output}`);\n\n    return output;\n  }\n}\n"]}