🔶 Extract v2.5：基于 LlamaParse 重构文档提取 Agent

今天，我们推出 Extract v2.5——新一代基于 schema 的文档提取 Agent。本次发布为所有层级带来准确率提升，并为最高层级 Agentic 和 Agentic Plus 增强了溯源能力。每页定价保持不变，因此无论你使用哪个层级，都能获得更高的每美元性能。

在 ExtractBench 上，三个层级的准确率均有所提升。Cost Effective 现在超过此前的 Agentic 层级，整体 value F1 从 87.1 提升至 93.9；Agentic 超过此前的 Agentic Plus，从 89.8 提升至 95.8。作为准确率最高的层级，Agentic Plus 从 95.1 提升至 96.4。

层级	Value F1 V2.0	Value F1 V2.5
Cost Effective	87.1	93.9
Agentic	89.8	95.8
Agentic Plus	95.1	96.4

平均每页成本（USD）：$0.00–$0.10

在底层架构方面，我们进行了多项改进，共同带来了上述性能提升，其中包括一个专为文档提取打造的全新 Agent harness。我们从最新的编程 Agent 中汲取灵感，并围绕模型进行了深度调优，以处理文档提取在视觉、推理和验证环节中的真实失败模式。最终得到的是一个专用 Agent：它能够交叉引用多页上下文，并将数值准确地锚定到具体来源。

首项变化是结构化推理（Structural Reasoning）。我们利用文档表示上的结构化推理，根据文档类型、版式和信息密度进行更高效、更有针对性的提取。Extract 会自适应地对更复杂的文档进行推理并投入更多计算，以实现最高准确率；对于复杂度较低的文档，它会以更低延迟处理，并将工作重点聚焦在必要之处。

本次发布为 Agentic 和 Agentic Plus 引入了高级引用（Advanced Citations），帮助定位提取数值的支持性证据。ExtractBench 的溯源分数中，Agentic 从 46.8 升至 80.6，Agentic Plus 从 46.4 升至 82.2；这意味着，即便面对最复杂的版式和提取任务，也能获得更多正确且贴合内容的边界框。

提取 API / LLM	分数
Sonnet 5.5	44.3
GPT-6 Sol	63.2
Opus 5.5	77.6
Reducto Deep Extract	77.5
Extend Max	50.8
Cost Effective V2.5	54.3
Agentic V2.5	80.6
Agentic Plus V2.5	82.2

平均每页成本（USD）：$0.00–$0.16

这些改进体现在 ExtractBench 的各类挑战任务中，包括长列表、跨页记录和扫描表单。以下展示了提取和溯源能力如何在这些任务中得到改善。

长列表

长列表可能导致即使是前沿 VLM 也会过早停止，或在重复记录中丢失位置。Extract v2.5 利用中间表示处理完整数据集，并根据你的 schema 验证输出结果。在长列表任务中，Cost Effective 的分数从 82.0 提升至 92.6，Agentic 从 86.1 提升至 95.5。

层级	长列表准确率 V2.0	长列表准确率 V2.5
Cost Effective	82.0	92.6
Agentic	86.1	95.5
Agentic Plus	94.5	96.3

在一份 17 页的基金申报文件中，Cost Effective 此前只能返回 238 项持仓中的 87 项。Extract v2.5 则返回全部 238 项，文档提取分数从 52.8 提升至 98.7。

跨页记录

在真实世界的文档中，提取必须跨页推理。某一页中的内容可能跨越多页，也可能需要与数页之后的附录关联。Extract v2.5 在构建输出时，更擅长保留字段之间的关系。

层级	跨页记录准确率 V2.0	跨页记录准确率 V2.5
Cost Effective	80.3	92.0
Agentic	85.5	96.5
Agentic Plus	93.6	96.7

在这份赠款计划表中，Agentic v2.0 在分页处停止，导致赠款用途和地址不完整。Extract v2.5 会将下一页的续接内容纳入同一条记录。

5

6

展开

address_line1

text 1800 WASHINGTON BLVD STE 340

city, state, zip_code

text "" → BALTIMORE MD 21230

purpose_of_grant

text EDUCATION, HUMAN & SOCIAL SERVICES, DISASTER RELIEF & RECOVERY, COMMUNITY DEVELOPMENT

扫描表单

扫描表单通常混合了印刷字段、叠加文本、手写内容和审核人员标注。这类表单中常见的扫描噪声，会使这些层次更加难以区分。印刷后的数值可能被划掉，并由手写内容更正。Extract v2.5 能够更好地结合上下文对字段进行视觉解读，并将所请求的数值与周围标记区分开来。现在，它也更擅长将提取数值溯源回原始文档，以便进行交叉核对和引用。

层级	扫描表单准确率 V2.0	扫描表单准确率 V2.5
Cost Effective	89.6	93.9
Agentic	90.9	95.7
Agentic Plus	94.4	96.1

在这份带有标注的许可证中，Agentic 此前将审核人员的日期与许可证编号合并，并将审核备注附加到淡水层深度中。Extract v2.5 会将原始数值和标注内容分离，并填入 schema 所要求的字段。

展开

amendment_permit_no

text "7/29/2019 13445" → "13445"

amendment_permit_no_review

text null → "7/29/2019"

deepest_freshwater_zone

text "4000 BUQW: 500(1800/3925) USDW: 4000 GEO ISOLATION: 4250" → "4000"

deepest_freshwater_zone_review

text "4000 BUQW: 500(1800/3925) USDW: 4000 GEO ISOLATION: 4250" → "BUQW: 500(1800/3925) USDW: 4000 GEO ISOLATION: 4250"

高级引用

高级引用可为高难度字段定位支持性证据的边界框，包括那些并未在文档中逐字出现的数值。此前，初版功能仅在 Agentic Plus 中提供。借助 v2.5，我们进一步提高了其溯源准确率，并将其扩展至 Agentic。

field_name Pleasanton (Edwards Lime) lease_name L.S.U. Leal current_operator_name SNG Operating, LLC operator_p5_no 797970 oil_lse_gas_id_no 14068 county Atascosa operator_address 1714 Fortview Road, #106 Austin, Texas 78704 well_nos 1 classification_oil false classification_gas false classification_other false effective_date May 1, 2010 change_operator true change_oil_condensate_gatherer false change_gas_gatherer false change_gas_purchaser false change_gas_purchaser_system_code false change_field_name false change_lease_name false new_rrc_oil_lease false new_rrc_gas_well false new_rrc_other_well false due_new_completion_recompletion false due_reclass_gas_to_oil false due_reclass_oil_to_gas false due_consolidation false

边界框引用可与置信度分数很好地配合使用。置信度分数帮助团队判断哪些数值可以自动进入后续流程，哪些需要进一步检查。随后，审核人员可在人工参与（human-in-the-loop）工作流中，利用引用将这些数值与原始文档进行核验。我们的置信度分数指南说明了如何在你自己的文档上评估这一能力。

我们如何构建 v2.5

在 LlamaIndex，我们深刻理解文档。我们已将这项专业能力构建为提取 Agent 可调用的工具，用于读取、转换和定位文档中的信息。这些工具依托专门打造的模型和文档处理方法，可应对不一致的文件格式、含糊的版式以及源坐标。

阅读顺序、表格结构，以及文本与其页面位置之间的关系，都会影响提取结果。我们的工具让 Agent 能够处理这些结构，并在构建输出时返回原始来源进行核验。

针对 v2.5，我们将此前支持 Agentic Plus 的 harness 引入 Cost Effective 和 Agentic，并根据各层级的成本限制调整其工具和提取策略。我们改进了任务组织方式、提供给 Agent 的上下文，以及 Agent 何时应更仔细检查文档。

要正确做出这些选择，需要进行广泛测试。一种适用于简短表单的方法，在处理数百条重复记录时可能表现不佳。我们针对这些任务评估了文档表示、工具和模型配置，并在准确率增益与其所需的时间和处理资源之间进行权衡。

我们还改进了 Agent 遵循你的 schema 和提取指令的方式，其中包括最多包含 3,200 个字段的任务。例如，如果记录 ID 对于将提取记录与你的数据库进行匹配至关重要，请在 prompt 中明确指出，以便 Agent 对其给予额外关注。

我们还新增了原生电子表格提取功能。在电子表格模式下，Agent 直接处理工作簿单元格，而不是处理扁平化表示，从而获得另一种处理结构化数据的方式。你可以在提取配置中启用它。

在你的文档上试用 v2.5

请使用现有 schema，在能够代表你工作流的文档上试用 v2.5。我们认为，这一版本将显著提升提取质量，尤其适用于此前需要人工清理、或可靠性不足以实现自动化的文档。我们期待看到你构建出的成果。

Today we’re introducing Extract v2.5, a new generation of our schema-based document extraction agents. This release brings accuracy improvements across all tiers as well as enhanced grounding capabilities to our highest tiers, Agentic and Agentic Plus. These improvements come with no increase in per-page pricing, ensuring higher performance per dollar no matter which tier you are using.

On ExtractBench, accuracy improved across all three tiers. Cost Effective now outperforms the previous Agentic tier, rising from 87.1 to 93.9 overall value F1, and Agentic outperforms the prior Agentic Plus, improving from 89.8 to 95.8. Agentic Plus, our highest-accuracy tier, rises from 95.1 to 96.4.

VALUE F1 V2.0V2.5 85 90 95 87.1 89.8 95.1 COST EFFECTIVE 93.9 AGENTIC 95.8 AGENTIC PLUS 96.4 $0.00$0.02$0.04$0.06$0.08$0.10 AVERAGE COST PER PAGE (USD)

Under the hood, we made architectural enhancements that contribute to these performance improvements, including a new agent harness, purpose-built for document extraction. We took inspiration from the latest coding agents, and hyper-tuned everything around the models to handle the real failure modes in document extraction, across vision, reasoning, and verification. The result is a specialized agent that can cross-reference context from multiple pages, and ground values in exact sources.

The first change is Structural Reasoning. We use structural reasoning**** over document representations for more efficient and tailored extraction based on document type, layout, and information density. Extract will adaptively reason and spend effort for your more complex docs to deliver maximum accuracy. For the less complex docs, it’ll process with lower latency and target its effort.

This release introduces Advanced Citations for Agentic and Agentic Plus, helping locate supporting evidence for extracted values. ExtractBench grounding scores rise from 46.8 to 80.6 for Agentic and from 46.4 to 82.2 for Agentic Plus, which means more correct and well fit bounding boxes for even your most complex layouts and extractions.

SCORE EXTRACTION APISLLMSV2.0V2.5 0 20 40 60 80 44.3 46.8 46.4 SONNET 5.5 63.2GPT-6 SOL 77.6OPUS 5.5 77.5REDUCTO DEEP EXTRACT 50.8EXTEND MAX 21.8 COST EFFECTIVE 54.3 AGENTIC 80.6 AGENTIC PLUS 82.2 $0.00$0.02$0.04$0.06$0.08$0.10$0.12$0.14$0.16 AVERAGE COST PER PAGE (USD)

These improvements show up across ExtractBench challenges, including long lists, records spanning pages, and scanned forms. The following show some examples of how extraction and grounding improve on these tasks.

Long lists

Long lists can cause even frontier VLMs to stop early or lose track of repeated records. Extract v2.5 uses intermediate representations to work with the full dataset and validate its output against your schema. Across long-list tasks, Cost Effective’s score rises from 82.0 to 92.6, and Agentic’s from 86.1 to 95.5.

LONG LIST ACCURACY V2.0V2.5 COST EFFECTIVE 82.0 92.6AGENTIC 86.1 95.5AGENTIC PLUS 94.5 96.3 7580859095100 SCORE

On a 17-page fund filing, Cost Effective previously returned 87 of 238 holdings. Extract v2.5 returns all 238, with the document’s extraction score rising from 52.8 to 98.7.

Records spanning pages

In real world documents, extraction must reason across pages. Content on one page might span multiple pages, or be joined with an appendix many pages later. Extract v2.5 is better at preserving the relationship between the fields as it builds the output.

CROSS-PAGE RECORDS ACCURACY V2.0V2.5 COST EFFECTIVE 80.3 92.0AGENTIC 85.5 96.5AGENTIC PLUS 93.6 96.7 7580859095100 SCORE

In this grant schedule, Agentic v2.0 stopped at the page break, leaving the grant’s purpose and address incomplete. Extract v2.5 includes the continuation from the next page in the same record.

5

6

Expand

address_line1 1800 WASHINGTON BLVD STE 340

city, state, zip_code "" → BALTIMORE MD 21230

purpose_of_grant EDUCATION, HUMAN & SOCIAL SERVICES, DISASTER RELIEF & RECOVERY, COMMUNITY DEVELOPMENT

×

Scanned forms

Scanned forms often mix printed fields, overlaid text, handwriting, and reviewer annotations. Scan noise common in such forms can make these layers even harder to distinguish. A printed value might be crossed out and hand corrected after printing. Extract v2.5 is better at visually interpreting the field in context and separating the requested value from the surrounding marks. It’s now also much better at grounding the extracted value back to the source document for cross checking and reference.

SCANNED FORMS ACCURACY V2.0V2.5 COST EFFECTIVE 89.6 93.9AGENTIC 90.9 95.7AGENTIC PLUS 94.4 96.1 7580859095100 SCORE

On this annotated permit, Agentic previously combined a reviewer’s date with the permit number and appended a reviewer note to the freshwater depth. Extract v2.5 separates the original values from the annotations into the fields requested by the schema.

Expand

amendment_permit_no "7/29/2019 13445" → "13445"

amendment_permit_no_review null → "7/29/2019"

deepest_freshwater_zone "4000 BUQW: 500(1800/3925) USDW: 4000 GEO ISOLATION: 4250" → "4000"

deepest_freshwater_zone_review "4000 BUQW: 500(1800/3925) USDW: 4000 GEO ISOLATION: 4250" → "BUQW: 500(1800/3925) USDW: 4000 GEO ISOLATION: 4250"

×

Advanced Citations

Advanced Citations locate bounding boxes of supporting evidence for difficult fields, including values that don’t appear word for word in the document. An initial version was previously available in Agentic Plus. With v2.5, we’ve improved its grounding accuracy further and extended it to Agentic.

field_name Pleasanton (Edwards Lime) lease_name L.S.U. Leal current_operator_name SNG Operating, LLC operator_p5_no 797970 oil_lse_gas_id_no 14068 county Atascosa operator_address 1714 Fortview Road, #106 Austin, Texas 78704 well_nos 1 classification_oil false classification_gas false classification_other false effective_date May 1, 2010 change_operator true change_oil_condensate_gatherer false change_gas_gatherer false change_gas_purchaser false change_gas_purchaser_system_code false change_field_name false change_lease_name false new_rrc_oil_lease false new_rrc_gas_well false new_rrc_other_well false due_new_completion_recompletion false due_reclass_gas_to_oil false due_reclass_oil_to_gas false due_consolidation false

Bounding box citations combine well with confidence scores. Confidence scores help teams decide which values can proceed automatically and which need a closer look. Reviewers can then use citations to check those values against the original document in human-in-the-loop workflows. Our confidence-score guide explains how to evaluate this on your own documents.

How we built v2.5

At LlamaIndex, we know documents deeply. We’ve built that expertise into tools our extraction agents can call to read, transform, and locate information in documents. Those tools draw on purpose-built models and document processing methods to handle inconsistent file formats, ambiguous layouts, and source coordinates.

Reading order, table structure, and the relationship between text and its position on the page all affect extraction. Our tools give agents ways to work with these structures and return to the source as they build the output.

For v2.5, we brought the harness behind Agentic Plus to Cost Effective and Agentic, adapting its tools and extraction strategies to each tier’s cost constraints. We refined how to organize the task, what context to give the agent, and when it should inspect the document more closely.

Getting those choices right took extensive testing. An approach that works well for a short form may struggle with hundreds of repeated records. We evaluated document representations, tools, and model configurations across these tasks, weighing accuracy gains against the time and processing they require.

We also worked on how agents follow your schema and extraction instructions, including tasks with up to 3,200 fields. If record IDs are essential for matching extracted records to your database, for example, call that out in your prompt so the agent can give them extra attention.

We’ve also added native spreadsheet extraction. In spreadsheet mode, agents work directly with workbook cells rather than a flattened representation, giving them another way to work with structured data. You can enable it in your extraction configuration.

Try v2.5 on your documents

Try v2.5 on documents representative of your workflow, using your existing schemas. We think this version will be a significant step up in extraction quality, especially for documents that previously needed manual cleanup or weren’t reliable enough to automate. We’re looking forward to seeing what you build.