aiex-cli 0.0.5-beta.3 → 0.0.5-beta.4

package/dist/cli.mjs

@@ -1,4 +1,4 @@
-import { A as doctorDiagnosticsTableRows, C as createConfig, D as package_default, E as name, O as version, S as AIConfigSchema, T as description, _ as DEFAULT_MINERU_API_CONFIG, a as parseJsonSchema, b as PLACEHOLDER_SCHEMA, c as recognizeImageText, d as t, f as getDefaultAIConfig, g as DEFAULT_MARKITDOWN_CONFIG, h as DEFAULT_MARKER_CONFIG, i as JsonSchemaDefinitionSchema, j as formatDoctorDiagnosticsJson, l as shouldUseImageOcrFallback, m as writeAIConfig, n as createMigrationConfig, o as toSnakeCase, p as readAIConfig, s as generateDrizzleSchema, t as collectDoctorDiagnostics, u as initI18n, v as DEFAULT_MINERU_CONFIG, w as seedConfig, x as PLACEHOLDER_TEXT, y as DEFAULT_PROMPT_CONFIG } from "./doctor-collector-CQPDBVTw.mjs";
+import { A as doctorDiagnosticsTableRows, C as createConfig, D as package_default, E as name, O as version, S as AIConfigSchema, T as description, _ as DEFAULT_MINERU_API_CONFIG, a as parseJsonSchema, b as PLACEHOLDER_SCHEMA, c as recognizeImageText, d as t, f as getDefaultAIConfig, g as DEFAULT_MARKITDOWN_CONFIG, h as DEFAULT_MARKER_CONFIG, i as JsonSchemaDefinitionSchema, j as formatDoctorDiagnosticsJson, l as shouldUseImageOcrFallback, m as writeAIConfig, n as createMigrationConfig, o as toSnakeCase, p as readAIConfig, s as generateDrizzleSchema, t as collectDoctorDiagnostics, u as initI18n, v as DEFAULT_MINERU_CONFIG, w as seedConfig, x as PLACEHOLDER_TEXT, y as DEFAULT_PROMPT_CONFIG } from "./doctor-collector-Cv7RArla.mjs";
 import { createRequire } from "node:module";
 import fs from "node:fs/promises";
 import os from "node:os";
@@ -18,7 +18,7 @@ import pc from "picocolors";
 import { Buffer } from "node:buffer";
 import * as XLSX from "xlsx";
 import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
-import { APICallError, Output, generateText, jsonSchema, tool } from "ai";
+import { APICallError, Output, generateText, jsonSchema } from "ai";
 import pRetry from "p-retry";
 import mime from "mime";
 import { jsonrepair } from "jsonrepair";
@@ -13565,12 +13565,13 @@ const HEADING_RE = /^(#{1,6})\s+(\S.*)$/;
 * Keeps tables and list blocks intact by splitting along paragraphs (\n\n)
 * when a section exceeds the maxSize limit.
 */
-function splitMarkdown(text$1, maxSize = 4e4) {
+function splitMarkdown(text$1, maxSize = 4e4, overlapSize = 0) {
 	const lines = text$1.split("\n");
 	const chunks = [];
 	let currentHeadings = [];
 	let currentChunkLines = [];
 	let currentSize = 0;
+	let hasNewLines = false;
 	const getMetadata = (headings) => {
 		return {
 			h1: headings[0] || void 0,
@@ -13579,9 +13580,15 @@ function splitMarkdown(text$1, maxSize = 4e4) {
 			h4: headings[3] || void 0
 		};
 	};
-	const flushChunk = () => {
-		if (currentChunkLines.length === 0) return;
+	const flushChunk = (isHeadingChange = false) => {
+		if (currentChunkLines.length === 0 || !hasNewLines) {
+			currentChunkLines = [];
+			currentSize = 0;
+			hasNewLines = false;
+			return;
+		}
 		const pageContent = currentChunkLines.join("\n");
+		let lastChunkContent = "";
 		if (pageContent.length > maxSize) {
 			const paragraphs = pageContent.split("\n\n");
 			let subLines = [];
@@ -13589,31 +13596,63 @@ function splitMarkdown(text$1, maxSize = 4e4) {
 			for (const para of paragraphs) {
 				const paraSize = para.length;
 				if (subSize + paraSize > maxSize && subLines.length > 0) {
+					const content = subLines.join("\n\n");
 					chunks.push({
-						pageContent: subLines.join("\n\n"),
+						pageContent: content,
 						metadata: getMetadata(currentHeadings)
 					});
-					subLines = [];
-					subSize = 0;
+					const overlapParas = [];
+					let currentOverlapSize = 0;
+					for (let j = subLines.length - 1; j >= 0; j--) {
+						const p = subLines[j];
+						if (currentOverlapSize + p.length > overlapSize && overlapParas.length > 0) break;
+						overlapParas.unshift(p);
+						currentOverlapSize += p.length + 2;
+					}
+					subLines = [...overlapParas];
+					subSize = currentOverlapSize;
 				}
 				subLines.push(para);
 				subSize += paraSize + 2;
 			}
-			if (subLines.length > 0) chunks.push({
-				pageContent: subLines.join("\n\n"),
+			if (subLines.length > 0) {
+				const content = subLines.join("\n\n");
+				chunks.push({
+					pageContent: content,
+					metadata: getMetadata(currentHeadings)
+				});
+				lastChunkContent = content;
+			}
+		} else {
+			chunks.push({
+				pageContent,
 				metadata: getMetadata(currentHeadings)
 			});
-		} else chunks.push({
-			pageContent,
-			metadata: getMetadata(currentHeadings)
-		});
-		currentChunkLines = [];
-		currentSize = 0;
+			lastChunkContent = pageContent;
+		}
+		if (!isHeadingChange && lastChunkContent && overlapSize > 0) {
+			const paragraphs = lastChunkContent.split("\n\n");
+			const overlapParas = [];
+			let currentOverlapSize = 0;
+			for (let j = paragraphs.length - 1; j >= 0; j--) {
+				const p = paragraphs[j];
+				if (currentOverlapSize + p.length > overlapSize && overlapParas.length > 0) break;
+				overlapParas.unshift(p);
+				currentOverlapSize += p.length + 2;
+			}
+			const overlapText = overlapParas.join("\n\n");
+			currentChunkLines = overlapText.split("\n");
+			currentSize = overlapText.length;
+		} else {
+			currentChunkLines = [];
+			currentSize = 0;
+		}
+		hasNewLines = false;
 	};
 	for (const line of lines) {
 		const headingMatch = line.match(HEADING_RE);
 		if (headingMatch) {
-			flushChunk();
+			flushChunk(true);
 			const depth = headingMatch[1].length;
 			const title = headingMatch[2].trim();
 			currentHeadings = currentHeadings.slice(0, depth - 1);
@@ -13621,220 +13660,13 @@ function splitMarkdown(text$1, maxSize = 4e4) {
 		}
 		currentChunkLines.push(line);
 		currentSize += line.length + 1;
-		if (currentSize > maxSize) flushChunk();
+		hasNewLines = true;
+		if (currentSize > maxSize) flushChunk(false);
 	}
-	flushChunk();
+	flushChunk(true);
 	return chunks;
 }
 
-//#endregion
-//#region src/core/ai-extraction/react-agent.ts
-async function extractStructuredDataWithAgent(input) {
-	const { config, schema, text: text$1, aiexDir, modelOverride, onAgentStep } = input;
-	if (!config.provider.apiKey) return {
-		success: false,
-		error: t("errors.ai.apiKeyMissing")
-	};
-	const chunks = splitMarkdown(text$1, 15e3);
-	const inputTokens = Math.ceil(text$1.length / 2);
-	const fieldCount = schema.properties ? Object.keys(schema.properties).length : 0;
-	const outputTokens = fieldCount > 0 ? fieldCount * 80 : void 0;
-	let selected;
-	try {
-		selected = modelOverride ?? selectModel({
-			models: config.provider.models,
-			isImage: false,
-			inputTokens,
-			outputTokens
-		});
-	} catch (e) {
-		return {
-			success: false,
-			error: e.message
-		};
-	}
-	const useTelemetry = !!(config.langfuse?.publicKey && config.langfuse.secretKey);
-	try {
-		if (useTelemetry) initLangfuse(config);
-		const provider = createOpenAICompatible({
-			baseURL: config.provider.baseURL,
-			name: "openai-compatible",
-			apiKey: config.provider.apiKey,
-			supportsStructuredOutputs: false
-		});
-		let finalExtractedData = null;
-		const tools = {
-			listChunks: tool({
-				description: "Get a list of all text chunks in the document, showing their chunk index ID, character size, and markdown heading hierarchy (metadata). Use this as a Table of Contents to locate sections of interest.",
-				parameters: z.object({}),
-				execute: async () => {
-					return chunks.map((c, idx) => ({
-						id: idx + 1,
-						size: c.pageContent.length,
-						headings: c.metadata
-					}));
-				}
-			}),
-			readChunk: tool({
-				description: "Read the full text content of a specific chunk by its ID.",
-				parameters: z.object({ chunkId: z.number().int().describe("The ID (1-based index) of the chunk to read.") }),
-				execute: async ({ chunkId }) => {
-					const index = chunkId - 1;
-					if (index < 0 || index >= chunks.length) return { error: `Invalid chunkId: ${chunkId}. Valid IDs are 1 to ${chunks.length}.` };
-					const chunk = chunks[index];
-					const headings = [];
-					if (chunk.metadata) {
-						if (chunk.metadata.h1) headings.push(chunk.metadata.h1);
-						if (chunk.metadata.h2) headings.push(chunk.metadata.h2);
-						if (chunk.metadata.h3) headings.push(chunk.metadata.h3);
-						if (chunk.metadata.h4) headings.push(chunk.metadata.h4);
-					}
-					return {
-						chunkId,
-						headings: headings.join(" > "),
-						content: chunk.pageContent
-					};
-				}
-			}),
-			searchChunks: tool({
-				description: "Search all chunks in the document for specific keywords or search terms. Returns matching chunk IDs and small matching context snippets.",
-				parameters: z.object({ query: z.string().describe("The keyword or search phrase to search for.") }),
-				execute: async ({ query }) => {
-					const results = [];
-					const lowercaseQuery = query.toLowerCase();
-					for (let i = 0; i < chunks.length; i++) {
-						const chunkText = chunks[i].pageContent;
-						const idx = chunkText.toLowerCase().indexOf(lowercaseQuery);
-						if (idx !== -1) {
-							const start = Math.max(0, idx - 60);
-							const end = Math.min(chunkText.length, idx + lowercaseQuery.length + 60);
-							const snippet = `...${chunkText.slice(start, end).replace(/\n/g, " ")}...`;
-							results.push({
-								chunkId: i + 1,
-								headings: chunks[i].metadata,
-								snippet
-							});
-						}
-					}
-					return results.slice(0, 10);
-				}
-			}),
-			submitExtraction: tool({
-				description: "Submit the final extracted JSON object conforming to the schema definition. Call this ONLY after you have gathered all necessary information.",
-				parameters: z.object({ data: z.any().describe("The extracted JSON object conforming to the target schema.") }),
-				execute: async ({ data }) => {
-					finalExtractedData = data;
-					return {
-						status: "success",
-						message: "Data submitted successfully. The extraction is now complete."
-					};
-				}
-			})
-		};
-		const outputSchema = schemaToExtractionOutputSchema(schema);
-		const systemPrompt = `You are a precise data extraction agent. Your goal is to extract structured information from a document to populate the target JSON schema.
-    
-Target JSON Schema structure to populate:
-${JSON.stringify(outputSchema, null, 2)}
-
-You are equipped with tools to browse the document dynamically:
-1. First, call listChunks to understand the document layout and what sections exist.
-2. Based on the schema fields, call readChunk or searchChunks to locate and read relevant content.
-3. You can make multiple tool calls. Do not guess. Check the text carefully.
-4. Once you have located and read all the necessary information, call the submitExtraction tool with the fully extracted JSON object.
-5. After calling submitExtraction, you should stop.
-
-CRITICAL RULES:
-1. Extract data strictly conforming to the types and properties of the Target JSON Schema.
-2. If a field's value cannot be found in the document after thorough search, set it to null.
-3. Do not invent any values.
-4. Call submitExtraction exactly once with the final JSON result.`;
-		const timeoutMs = (config.provider.timeout ?? 300) * 1e3;
-		const result = await generateText({
-			model: provider.chatModel(selected.name),
-			system: systemPrompt,
-			prompt: "Please start by listing the chunks to understand the document structure, then gather the required facts and submit the final JSON extraction.",
-			tools,
-			maxSteps: 12,
-			abortSignal: AbortSignal.timeout(timeoutMs),
-			experimental_telemetry: { isEnabled: useTelemetry },
-			onStepFinish({ text: text$2, toolCalls }) {
-				if (onAgentStep) onAgentStep({
-					thought: text$2,
-					toolCalls
-				});
-			}
-		});
-		if (!finalExtractedData) {
-			if (result.text) try {
-				finalExtractedData = safeParseJSON(result.text);
-			} catch {}
-		}
-		if (!finalExtractedData) return {
-			success: false,
-			error: "Agent finished without submitting structured data."
-		};
-		const validation = validateExtractedData(schema, finalExtractedData);
-		if (!validation.success) {
-			const correctionSystemPrompt = `You are a precise data correction assistant. Your task is to correct validation errors in a previously generated JSON object to make it comply with the JSON Schema.
-      
-JSON Schema Definition:
-${JSON.stringify(outputSchema, null, 2)}
-
-Validation Errors:
-${validation.error}
-
-Original Incorrect JSON:
-${JSON.stringify(finalExtractedData, null, 2)}
-
-Please output the corrected JSON object. Return ONLY the corrected JSON object, with no markdown tags or explanations.`;
-			const correctedData = safeParseJSON((await generateText({
-				model: provider.chatModel(selected.name),
-				system: correctionSystemPrompt,
-				prompt: "Please correct the JSON output now.",
-				abortSignal: AbortSignal.timeout(timeoutMs),
-				experimental_telemetry: { isEnabled: useTelemetry }
-			})).text);
-			const secondValidation = validateExtractedData(schema, correctedData);
-			if (!secondValidation.success) return {
-				success: false,
-				error: `Agent output validation failed: ${secondValidation.error}`
-			};
-			finalExtractedData = correctedData;
-		}
-		const outputDir = path.resolve(aiexDir, config.extraction?.outputDir?.replace(".aiex/", "") ?? "extracted");
-		await fs.mkdir(outputDir, { recursive: true });
-		const timestamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
-		const outputFileName = `${schema.table.name}-${timestamp}.json`;
-		const outputPath = path.join(outputDir, outputFileName);
-		await writeFile(outputPath, finalExtractedData, {
-			spaces: 2,
-			EOL: "\n"
-		});
-		let totalPromptTokens = 0;
-		let totalCompletionTokens = 0;
-		if (result.usage) {
-			totalPromptTokens = result.usage.inputTokens ?? 0;
-			totalCompletionTokens = result.usage.outputTokens ?? 0;
-		}
-		return {
-			success: true,
-			outputPath,
-			data: finalExtractedData,
-			tokensUsed: {
-				prompt: totalPromptTokens,
-				completion: totalCompletionTokens,
-				total: totalPromptTokens + totalCompletionTokens
-			}
-		};
-	} catch (error) {
-		return {
-			success: false,
-			error: getErrorMessage(error)
-		};
-	}
-}
-
 //#endregion
 //#region src/core/extraction-audit.ts
 const AUDIT_ID_RE = /^[\w.-]+$/;
@@ -14719,7 +14551,6 @@ async function processOneFile(aiexDir, config, aiConfig, schemaName, filePath, m
 		modelOverride,
 		insert: options?.insert,
 		force: options?.force,
-		agent: options?.agent,
 		quiet: false
 	});
 	if (result.success) {
@@ -14759,8 +14590,7 @@ async function runBatchExtraction(aiexDir, config, aiConfig, schemaName, dir, gl
 		})}`);
 		if (await processOneFile(aiexDir, config, aiConfig, schemaName, file, modelOverride, {
 			insert: options?.insert,
-			force: options?.force,
-			agent: options?.agent
+			force: options?.force
 		})) successCount++;
 		else failCount++;
 	}
@@ -14779,6 +14609,42 @@ async function runBatchExtraction(aiexDir, config, aiConfig, schemaName, dir, gl
 //#endregion
 //#region src/core/extract-runner.ts
 const JSON_EXT_RE$1 = /\.json$/;
+async function limitConcurrency(concurrency, items, fn) {
+	const results = Array.from({ length: items.length });
+	let nextIndex = 0;
+	async function worker() {
+		while (nextIndex < items.length) {
+			const currentIndex = nextIndex++;
+			results[currentIndex] = await fn(items[currentIndex], currentIndex);
+		}
+	}
+	const workers = Array.from({ length: Math.min(concurrency, items.length) }, worker);
+	await Promise.all(workers);
+	return results;
+}
+function getSchemaKeywords(schema) {
+	const keywords = /* @__PURE__ */ new Set();
+	function walk(properties) {
+		if (!properties) return;
+		for (const [name$1, prop] of Object.entries(properties)) {
+			keywords.add(name$1.toLowerCase());
+			const parts = name$1.replace(/([a-z0-9])([A-Z])/g, "$1 $2").split(/[\s._:/\\-]+/g);
+			for (const part of parts) if (part.length > 1) keywords.add(part.toLowerCase());
+			if (prop && typeof prop === "object") {
+				const p = prop;
+				if (typeof p.title === "string") keywords.add(p.title.toLowerCase());
+				if (typeof p.description === "string") {
+					const descParts = p.description.toLowerCase().match(/[\p{L}\p{N}_-]+/gu) ?? [];
+					for (const d of descParts) if (d.length > 2) keywords.add(d);
+				}
+				if (p.type === "object") walk(p.properties);
+				if (p.type === "array" && p.items?.type === "object") walk(p.items.properties);
+			}
+		}
+	}
+	walk(schema.properties);
+	return Array.from(keywords);
+}
 async function ensureDatabaseReady(dbPath, schema) {
 	try {
 		await fs.access(dbPath);
@@ -14852,42 +14718,48 @@ async function extractSingle(aiexDir, config, aiConfig, schemaName, text$1, file
 	if (!options?.quiet) s.start(filePath ? t("command.extract.file.extractedFrom", { file: path.basename(filePath) }) : t("command.extract.file.extracting"));
 	const CHUNK_LIMIT = 4e4;
 	let result;
-	if (options?.agent || aiConfig.extraction?.mode === "react") {
-		if (!options?.quiet) consola.info(t("command.extract.file.reactAgentMode"));
-		const agentResult = await extractStructuredDataWithAgent({
-			config: aiConfig,
-			schema: schemaLoad.schema,
-			text: text$1 ?? "",
-			aiexDir,
-			modelOverride,
-			onAgentStep(step) {
-				if (!options?.quiet) {
-					if (step.thought) {
-						const thoughtPreview = step.thought.length > 100 ? `${step.thought.slice(0, 100)}...` : step.thought;
-						s.message(`${pc.cyan(t("command.extract.file.agentThought"))}: ${thoughtPreview.replace(/\n/g, " ")}`);
-					}
-					if (step.toolCalls && step.toolCalls.length > 0) for (const call of step.toolCalls) consola.info(`[Agent Action] Calling tool: ${pc.green(call.toolName)}`);
-				}
-			}
-		});
-		if (!agentResult.success) {
-			if (!options?.quiet) {
-				s.stop(t("command.extract.file.extractFail"));
-				consola.error(agentResult.error);
-			}
-			return {
-				success: false,
-				error: agentResult.error
-			};
-		}
-		result = agentResult;
-	} else if (text$1 && text$1.length > CHUNK_LIMIT) {
+	if (text$1 && text$1.length > CHUNK_LIMIT) {
 		if (!options?.quiet) consola.info(t("command.extract.file.chunking", {
 			length: text$1.length,
 			limit: CHUNK_LIMIT
 		}));
-		const finalDocs = splitMarkdown(text$1, CHUNK_LIMIT);
+		const finalDocs = splitMarkdown(text$1, CHUNK_LIMIT, aiConfig.extraction?.overlapSize ?? 2e3);
 		if (!options?.quiet) consola.info(t("command.extract.file.chunksCount", { count: finalDocs.length }));
+		let processedDocs = finalDocs;
+		if (!!aiConfig.extraction?.preFiltering && finalDocs.length > 1) {
+			const preFilteringLimit = aiConfig.extraction?.preFilteringLimit ?? 5;
+			const keywords = getSchemaKeywords(schemaLoad.schema);
+			const scoredChunks = finalDocs.map((doc, idx) => {
+				if (idx === 0) return {
+					index: idx,
+					score: Number.POSITIVE_INFINITY
+				};
+				let score = 0;
+				const docTextLower = doc.pageContent.toLowerCase();
+				for (const kw of keywords) {
+					let pos = docTextLower.indexOf(kw);
+					while (pos !== -1) {
+						score++;
+						pos = docTextLower.indexOf(kw, pos + kw.length);
+					}
+				}
+				return {
+					index: idx,
+					score
+				};
+			}).slice(1).sort((a, b) => b.score - a.score);
+			const selectedIndices = new Set([0]);
+			let keptCount = 0;
+			for (const sc of scoredChunks) if (sc.score > 0 && keptCount < preFilteringLimit) {
+				selectedIndices.add(sc.index);
+				keptCount++;
+			}
+			processedDocs = finalDocs.filter((_, idx) => selectedIndices.has(idx));
+			if (!options?.quiet) consola.info(t("command.extract.file.preFiltering", {
+				original: finalDocs.length,
+				filtered: processedDocs.length
+			}));
+		}
 		const chunkResults = [];
 		const accumulatedTokens = {
 			prompt: 0,
@@ -14896,53 +14768,68 @@ async function extractSingle(aiexDir, config, aiConfig, schemaName, text$1, file
 		};
 		let success = true;
 		let errorMsg = "";
-		for (let i = 0; i < finalDocs.length; i++) {
-			const doc = finalDocs[i];
-			if (!options?.quiet) s.message(t("command.extract.file.extractingChunk", {
-				current: i + 1,
-				total: finalDocs.length
-			}));
-			const headings = [];
-			if (doc.metadata) {
-				if (doc.metadata.h1) headings.push(doc.metadata.h1);
-				if (doc.metadata.h2) headings.push(doc.metadata.h2);
-				if (doc.metadata.h3) headings.push(doc.metadata.h3);
-				if (doc.metadata.h4) headings.push(doc.metadata.h4);
-			}
-			let chunkText = doc.pageContent;
-			if (headings.length > 0) chunkText = `> **[Context]** Belong to: ${headings.join(" > ")}\n\n${chunkText}`;
-			const chunkResult = await extractStructuredData({
-				config: aiConfig,
-				schema: schemaLoad.schema,
-				text: chunkText,
-				aiexDir,
-				modelOverride,
-				onRetry(info) {
-					if (!options?.quiet) s.message(t("command.extract.file.extractRetryChunk", {
-						current: i + 1,
-						total: finalDocs.length,
-						code: info.statusCode,
-						delay: info.delayMs / 1e3,
-						attempt: info.attempt,
-						max: info.maxRetries
-					}));
+		const extractionTasks = processedDocs.map((doc, i) => {
+			return async () => {
+				if (!success) return;
+				const headings = [];
+				if (doc.metadata) {
+					if (doc.metadata.h1) headings.push(doc.metadata.h1);
+					if (doc.metadata.h2) headings.push(doc.metadata.h2);
+					if (doc.metadata.h3) headings.push(doc.metadata.h3);
+					if (doc.metadata.h4) headings.push(doc.metadata.h4);
 				}
-			});
-			if (!chunkResult.success) {
-				success = false;
-				errorMsg = chunkResult.error || t("common.unknownError");
-				if (!options?.quiet) {
-					s.stop(t("command.extract.file.extractFailChunk", { current: i + 1 }));
-					consola.error(errorMsg);
+				let chunkText = doc.pageContent;
+				if (headings.length > 0) chunkText = `> **[Context]** Belong to: ${headings.join(" > ")}\n\n${chunkText}`;
+				const chunkResult = await extractStructuredData({
+					config: aiConfig,
+					schema: schemaLoad.schema,
+					text: chunkText,
+					aiexDir,
+					modelOverride,
+					onRetry(info) {
+						if (!options?.quiet) s.message(t("command.extract.file.extractRetryChunk", {
+							current: i + 1,
+							total: processedDocs.length,
+							code: info.statusCode,
+							delay: info.delayMs / 1e3,
+							attempt: info.attempt,
+							max: info.maxRetries
+						}));
+					}
+				});
+				if (!chunkResult.success) {
+					success = false;
+					errorMsg = chunkResult.error || t("common.unknownError");
+					if (!options?.quiet) {
+						s.stop(t("command.extract.file.extractFailChunk", { current: i + 1 }));
+						consola.error(errorMsg);
+					}
+					return;
 				}
-				break;
-			}
-			if (chunkResult.data) chunkResults.push(chunkResult.data);
-			if (chunkResult.tokensUsed) {
-				accumulatedTokens.prompt += chunkResult.tokensUsed.prompt ?? 0;
-				accumulatedTokens.completion += chunkResult.tokensUsed.completion ?? 0;
-				accumulatedTokens.total += chunkResult.tokensUsed.total ?? 0;
-			}
+				if (chunkResult.data) chunkResults.push(chunkResult.data);
+				if (chunkResult.tokensUsed) {
+					accumulatedTokens.prompt += chunkResult.tokensUsed.prompt ?? 0;
+					accumulatedTokens.completion += chunkResult.tokensUsed.completion ?? 0;
+					accumulatedTokens.total += chunkResult.tokensUsed.total ?? 0;
+				}
+			};
+		});
+		const concurrency = Math.min(aiConfig.extraction?.concurrency ?? 2, 2);
+		if (!options?.quiet && processedDocs.length > 0) s.message(t("command.extract.file.extractingChunk", {
+			current: 1,
+			total: processedDocs.length
+		}));
+		try {
+			await limitConcurrency(concurrency, extractionTasks, async (task, idx) => {
+				if (!options?.quiet && success) s.message(t("command.extract.file.extractingChunk", {
+					current: idx + 1,
+					total: processedDocs.length
+				}));
+				await task();
+			});
+		} catch (e) {
+			success = false;
+			errorMsg = e instanceof Error ? e.message : String(e);
 		}
 		if (!success) return {
 			success: false,
@@ -15001,6 +14888,11 @@ async function extractSingle(aiexDir, config, aiConfig, schemaName, text$1, file
 	}
 	if (!options?.quiet) s.stop(t("command.extract.file.extractComplete"));
 	if (result.outputPath && !options?.quiet) consola.success(t("command.extract.file.resultSaved", { path: pc.cyan(result.outputPath) }));
+	if (result.evidenceSummary && !options?.quiet) {
+		const summary = result.evidenceSummary;
+		const issueText = summary.issueCount > 0 ? pc.yellow(String(summary.issueCount)) : pc.green("0");
+		consola.info(pc.gray(`Evidence coverage: ${summary.evidenceCount}/${summary.fieldCount} fields, found ${summary.foundCount}, inferred ${summary.inferredCount}, missing ${summary.missingCount}, issues ${issueText}`));
+	}
 	if (result.tokensUsed && !options?.quiet) consola.info(pc.gray(t("command.extract.file.tokenUsage", {
 		prompt: result.tokensUsed.prompt,
 		completion: result.tokensUsed.completion,
@@ -15029,6 +14921,7 @@ async function extractSingle(aiexDir, config, aiConfig, schemaName, text$1, file
 						outputPath: result.outputPath,
 						data: result.data,
 						tablesInserted: insertResult.tablesInserted,
+						evidenceSummary: result.evidenceSummary,
 						tokensUsed: result.tokensUsed
 					};
 				} else {
@@ -15055,11 +14948,12 @@ async function extractSingle(aiexDir, config, aiConfig, schemaName, text$1, file
 		success: true,
 		outputPath: result.outputPath,
 		data: result.data,
+		evidenceSummary: result.evidenceSummary,
 		tokensUsed: result.tokensUsed
 	};
 }
 async function runAuditedExtraction(options) {
-	const { aiexDir, config, aiConfig, schemaName, source, modelOverride, retryOf, insert, force, quiet = false, agent = false } = options;
+	const { aiexDir, config, aiConfig, schemaName, source, modelOverride, retryOf, insert, force, quiet = false } = options;
 	let fileHash;
 	let isPlainTextFile = false;
 	if (source.type === "file") {
@@ -15127,8 +15021,7 @@ async function runAuditedExtraction(options) {
 		} else text$1 = source.text;
 		const r = await extractSingle(aiexDir, config, aiConfig, schemaName, text$1, filePath, modelOverride, {
 			quiet,
-			insert,
-			agent
+			insert
 		});
 		if (r.success) {
 			let notionPages;
@@ -15168,6 +15061,7 @@ async function runAuditedExtraction(options) {
 				outputName: updated.outputName,
 				tablesInserted: updated.tablesInserted,
 				notionPages: updated.notionPages,
+				evidenceSummary: r.evidenceSummary,
 				tokensUsed: updated.tokensUsed,
 				auditId: updated.id,
 				fileHash
@@ -15587,12 +15481,6 @@ const extractCommand = defineCommand({
 			type: "boolean",
 			description: t("command.extract.args.force"),
 			default: false
-		},
-		agent: {
-			type: "boolean",
-			alias: "a",
-			description: "Enable ReAct agent extraction mode",
-			default: false
 		}
 	},
 	async run({ args, rawArgs }) {
@@ -15620,8 +15508,7 @@ const extractCommand = defineCommand({
 			}
 			const result$1 = await runBatchExtraction(aiexDir, config, aiConfig, args.schema, args.dir, args.glob, modelOverride, {
 				insert: !args.noInsert,
-				force: args.force,
-				agent: args.agent
+				force: args.force
 			});
 			if (!result$1.ok) {
 				failCommand(result$1.error);
@@ -15652,8 +15539,7 @@ const extractCommand = defineCommand({
 			modelOverride,
 			insert: !args.noInsert,
 			force: args.force,
-			quiet: false,
-			agent: args.agent
+			quiet: false
 		});
 		if (!result.success) {
 			failCommand(result.error);
@@ -16303,6 +16189,7 @@ function aiRoutes(config) {
 //#endregion
 //#region src/core/data-service.ts
 const FILE_REGEX = /\.json$/;
+const EVIDENCE_FILE_SUFFIX = ".evidence.json";
 const EXTRACTION_TIMESTAMP_RE = /-\d{4}-\d{2}-\d{2}T/;
 const INTERNAL_ROWID_COLUMN = "__aiex_rowid";
 const TIMESTAMP_CLEANUP = /(\d{2})-(\d{2})-(\d{2})/;
@@ -16318,6 +16205,24 @@ function getAuditNotionStatus(record) {
 	if (record.status === "failed") return "failed";
 	return "not_synced";
 }
+async function readEvidenceSummary(extractedDir, outputName) {
+	const evidencePath = path.join(extractedDir, outputName.replace(FILE_REGEX, EVIDENCE_FILE_SUFFIX));
+	try {
+		const coverage = (await readFile(evidencePath))?.coverage;
+		if (!coverage || typeof coverage !== "object") return void 0;
+		return {
+			path: evidencePath,
+			fieldCount: Number(coverage.fieldCount) || 0,
+			evidenceCount: Number(coverage.evidenceCount) || 0,
+			foundCount: Number(coverage.foundCount) || 0,
+			missingCount: Number(coverage.missingCount) || 0,
+			inferredCount: Number(coverage.inferredCount) || 0,
+			issueCount: Number(coverage.issueCount) || 0
+		};
+	} catch {
+		return;
+	}
+}
 async function getRowExtractionActions(aiexDir, tableName) {
 	const actions = /* @__PURE__ */ new Map();
 	const auditRecords = await listExtractionAuditRecords(aiexDir);
@@ -16345,7 +16250,7 @@ async function listExtractions(config) {
 	const aiexDir = path.dirname(config.schemaPath);
 	const extractedDir = path.join(aiexDir, "extracted");
 	await fs.mkdir(extractedDir, { recursive: true });
-	const jsonFiles = (await fs.readdir(extractedDir)).filter((f) => f.endsWith(".json") && !f.endsWith(".prompt.md"));
+	const jsonFiles = (await fs.readdir(extractedDir)).filter((f) => f.endsWith(".json") && !f.endsWith(".prompt.md") && !f.endsWith(EVIDENCE_FILE_SUFFIX));
 	const auditRecords = await listExtractionAuditRecords(aiexDir);
 	const auditByOutputName = new Map(auditRecords.map((record) => [record.outputName, record]));
 	const records = [];
@@ -16364,6 +16269,7 @@ async function listExtractions(config) {
 				timestamp,
 				fileSize: stat.size,
 				modifiedAt: stat.mtime.toISOString(),
+				evidenceSummary: await readEvidenceSummary(extractedDir, file),
 				notionStatus: notionPages ? "synced" : audit?.status === "failed" ? "failed" : "not_synced",
 				notionPages,
 				notionError: !notionPages && audit?.status === "failed" ? audit.error : void 0
@@ -16543,6 +16449,7 @@ async function retryNotionSync(config, fileName) {
 
 //#endregion
 //#region src/server/routes/data.ts
+const JSON_FILE_SUFFIX_RE = /\.json$/;
 const tableParamSchema = z.object({ name: z.string().regex(/^[a-z][a-z0-9_]*$/) });
 const extractionFileParamSchema = z.object({ name: z.string().regex(/^[\w.-]+\.json$/).refine((name$1) => name$1 === path.basename(name$1) && !name$1.includes("..")) });
 const tableQuerySchema = z.object({
@@ -16595,10 +16502,22 @@ function dataRoutes(config) {
 		const filePath = path.join(extractedDir, name$1);
 		try {
 			const content = await fs.readFile(filePath, "utf-8");
+			const evidencePath = path.join(extractedDir, name$1.replace(JSON_FILE_SUFFIX_RE, ".evidence.json"));
+			let evidenceSummary;
+			try {
+				const evidence = JSON.parse(await fs.readFile(evidencePath, "utf-8"));
+				evidenceSummary = evidence?.coverage ? {
+					...evidence.coverage,
+					path: evidencePath
+				} : void 0;
+			} catch {
+				evidenceSummary = void 0;
+			}
 			return c.json({
 				success: true,
 				content,
-				name: name$1
+				name: name$1,
+				evidenceSummary
 			});
 		} catch {
 			return c.json({ error: t("server.extractionNotFound") }, 404);
@@ -16742,6 +16661,7 @@ function extractRoutes(config) {
 				outputName: result.outputName,
 				tablesInserted: result.tablesInserted,
 				notionPages: result.notionPages,
+				evidenceSummary: result.evidenceSummary,
 				tokensUsed: result.tokensUsed,
 				auditId: result.auditId
 			}, 200);
@@ -16809,6 +16729,7 @@ function extractRoutes(config) {
 			outputName: result.outputName,
 			tablesInserted: result.tablesInserted,
 			notionPages: result.notionPages,
+			evidenceSummary: result.evidenceSummary,
 			tokensUsed: result.tokensUsed,
 			auditId: result.auditId
 		}, 200);