Skip to content

pipelines.md

Latest commit

 

History

History
451 lines (348 loc) · 29.6 KB

pipelines.md

File metadata and controls

451 lines (348 loc) · 29.6 KB
Raw

Available Pipelines

To see the full list:

uv run extract-bench pipelines

Extract Pipelines (ExtractBench)

Extraction pipelines run per split over the ExtractBench dataset:

uv run extract-bench download
uv run extract-bench run <pipeline_name>
Pipeline Provider Notes
llamaextract_cost_effective / llamaextract_agentic LlamaExtract V2 API Extract v2.6 with citations enabled (LLAMA_CLOUD_API_KEY)
llamaextract_agentic_plus LlamaExtract V2 API Extract v2.6, citations enabled
llamaextract_agentic_plus_max_effort LlamaExtract V2 API Agentic Plus with effort=max, latest, citations and confidence scores enabled
openai_gpt_5_4_extract_oneshot_structured_output_file (+ _nano) OpenAI Responses one-shot structured output over the uploaded file
openai_gpt_6_astra_reasoning_low_extract_oneshot_structured_output_file OpenAI Responses one-shot structured output over the uploaded file (reasoning: low)
gemini_3_5_flash_extract_oneshot_structured_output_file Gemini one-shot structured output (thinking: low)
gemini_3_6_flash_extract_oneshot_structured_output_file Gemini one-shot structured output (thinking: medium)
gemini_3_7_flash_extract_oneshot_structured_output_file Gemini one-shot structured output (thinking: medium)
gemini_3_8_flash_extract_oneshot_structured_output_file Gemini one-shot structured output (thinking: low)
anthropic_haiku_4_5_extract_oneshot_structured_output_file Anthropic one-shot structured output
glm_5_3_flash_extract_oneshot_structured_output_file GLM (z.ai) one-shot structured output over the file, json_object mode; uses GLM_ZAI_API_KEY
deepseek_v4_1_flash_extract_oneshot_structured_output_file DeepSeek one-shot vision (rasterized pages), json_object mode, thinking off; uses DEEPSEEK_API_KEY
*_extract_twostage_parse_agentic_structured_output_text OpenAI / Gemini / Anthropic LlamaParse agentic markdown → text extract; cost totals parse + extract
deepseek_v4_pro_extract_twostage_parse_agentic_structured_output_text DeepSeek (Fireworks) two-stage text extract
kimi_k3_extract_oneshot_structured_output_file Kimi K3 (Fireworks) one-shot vision (rasterized pages), json_object mode; uses FIREWORKS_API_KEY
gemini_cli_extract_gemini_4_argon_medium_thinking_vision_evidence Gemini CLI 0.59.0 Argon medium thinking with page/box citations; setup
claude_code_extract_opus_4_8 / claude_code_extract_opus_5_5 Claude Code CLI agentic extraction; cost from CLI total_cost_usd; Opus 5.5 needs CLI 2.1.280+; setup
codex_code_extract_gpt_5_4_low / codex_code_extract_gpt_5_5_low / codex_code_extract_gpt_5_5_high Codex CLI agentic extraction; cost estimated from token usage; setup
claude_code_extract_opus_4_8_evidence / claude_code_extract_opus_5_5_evidence Claude Code CLI agentic extraction with page and box citations in citations.json
codex_code_extract_gpt_5_5_low_evidence / codex_code_extract_gpt_5_6_sol_low_evidence / codex_code_extract_gpt_5_6_terra_low_evidence Codex CLI low-effort extraction with page and box citations; unrestricted shell
codex_code_extract_gpt_5_6_luna_medium_evidence Codex CLI medium-effort extraction with page and box citations; unrestricted shell, no cost cap
codex_code_extract_gpt_6_luna_medium_evidence / codex_code_extract_gpt_6_luna_high_evidence / codex_code_extract_gpt_6_luna_xhigh_evidence / codex_code_extract_gpt_6_sol_low_evidence / codex_code_extract_gpt_6_sol_medium_evidence / codex_code_extract_gpt_6_sol_high_evidence Codex CLI GPT-6 Luna and Sol extraction with page and box citations at the effort in the name; unrestricted shell, no cost cap
codex_code_extract_gpt_6_astra_xhigh_evidence / codex_code_extract_gpt_6_1_sol_{low,medium,high,xhigh}_evidence / codex_code_extract_gpt_6_1_terra_{low,medium}_evidence Codex CLI GPT-6 Astra, GPT-6.1 Sol and GPT-6.1 Terra extraction with page and box citations at the effort in the name; unrestricted shell, no cost cap
codex_code_extract_gpt_6_luna_{low,medium,high,xhigh}_evidence_prompt_opt Codex CLI GPT-6 Luna evidence extraction plus a short prompt addendum: always write the output files, the documents are authorized for extraction, no field needs web access
pi_extract_hy4_vision_evidence pi 0.84.1 Tencent Hy4 on TokenHub, reasoning high, vision, page/box citations; uses TENCENT_TOKENHUB_API_KEY; setup
reducto_extract / reducto_extract_v4 / reducto_deep_extract Reducto _v4 uses Reducto's v4 extract model
extend_extract Extend Performance + Review Agent latest stable model, Review Agent scores, line/property citations; async API 2026-02-09
extend_extract_max Extend Max Context + Review Agent same settings as Performance, with large_array_max_context
extend_extract_operator Extend Operator extraction on the 2026-02-09 API with word-level citations; cost from the run's reported usage (EXTEND_API_KEY)
landingai_extract LandingAI ADE
parseur_extract_v3_vision Parseur AI Vision engine v3; no citations (PARSEUR_API_KEY)
datalab_extract_turbo Datalab Turbo Image-only extraction; no parsing, citations, or confidence scores
datalab_extract_fast Datalab Fast Fast extraction with source citations; a separate, free confidence step runs afterward
datalab_extract_balanced Datalab Balanced Balanced extraction with source citations and inline field verification
datalab_extract_accurate Datalab Accurate Accurate extraction with source citations and inline field verification
pulse_schema_non_effort / pulse_schema_effort Pulse Pulse Ultra 2, WLBB enabled, then /schema with effort off/on (PULSE_API_KEY)
lift_extract Self-hosted lift SDK requires LIFT_ENDPOINT_URL
qwen3_6_35b_a3b_fp8_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Qwen3.6-35B-A3B-FP8, json_object mode; requires QWEN35_SERVER_URL
gemma4_26b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Gemma-4-26B-A4B, json_object mode; requires GEMMA4_SERVER_URL
nuextract_2_8b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM NuExtract-2.0-8B, json_object mode; requires NUEXTRACT2_SERVER_URL
internvl3_5_8b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM InternVL3.5-8B, json_object mode; requires INTERNVL35_8B_SERVER_URL
internvl3_5_14b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM InternVL3.5-14B, json_object mode; requires INTERNVL35_14B_SERVER_URL
minicpm_v_4_5_vllm_extract_oneshot_structured_output_file Self-hosted vLLM MiniCPM-V-4.5, json_object mode; requires MINICPM_V_SERVER_URL
qwen3_5_35b_a3b_fp8_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Qwen3.5-35B-A3B-FP8, json_object mode; requires QWEN3_5_35B_SERVER_URL
qwen3_5_9b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Qwen3.5-9B, json_object mode; requires QWEN3_5_9B_SERVER_URL
gemma4_e4b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Gemma-4-E4B, json_object mode; requires GEMMA4_E4B_SERVER_URL
gemma4_e2b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Gemma-4-E2B, json_object mode; requires GEMMA4_E2B_SERVER_URL
kimi_vl_a3b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Kimi-VL-A3B-Instruct, json_object mode; requires KIMI_VL_SERVER_URL
glm_4_6v_flash_vllm_extract_oneshot_structured_output_file Self-hosted vLLM GLM-4.6V-Flash, json_object mode; requires GLM_4_6V_SERVER_URL
qwen3_8_27b_fp8_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Qwen3.8-27B-FP8, json_object mode; requires QWEN3_8_27B_SERVER_URL
qwen3_8_flash_next_fp8_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Qwen3.8-Flash-Next-FP8, json_object mode; requires QWEN3_8_FLASH_NEXT_SERVER_URL
qwen3_5_2b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Qwen3.5-2B, json_object mode; requires QWEN3_5_2B_SERVER_URL
qwen3_5_0_8b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Qwen3.5-0.8B, json_object mode; requires QWEN3_5_0_8B_SERVER_URL
gemma4_12b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Gemma-4-12B, json_object mode; requires GEMMA4_12B_SERVER_URL
qwen3_5_4b_vllm_extract_oneshot_structured_output_file Self-hosted vLLM Qwen3.5-4B, json_object mode; requires QWEN3_5_4B_SERVER_URL
nuextract3_extract Self-hosted vLLM schema converted to a NuExtract template; requires NUEXTRACT3_SERVER_URL

Parse Pipelines

Parse pipelines (inherited from ParseBench, used here by the two-stage extract baselines and grounding cross-eval) can be run with:

uv run extract-bench run <pipeline_name>

Setup

Copy .env.example to .env and fill in the API keys / endpoints for the providers you want to use:

cp .env.example .env

Coding-agent setup

These pipelines use a separately installed CLI to inspect documents and write extraction results with local tools. Run them in a container or VM: they execute shell commands without interactive approval, and evidence pipelines can have unrestricted shell access.

Run the commands below from the repository root, with Node.js 22+ and npm installed. Replace X.Y.Z in each example with the exact CLI release to use. Each example runs the full public Pro dataset at concurrency 1; the benchmark downloads it automatically. For a single document, replace --dataset pro with --file /path/to/document.pdf, keeping its matching document.test.json beside it. Omit --dataset pro to use original ExtractBench.

Set harness_version in pipeline config, or pass --harness_version X.Y.Z to extract-bench run or extract-bench inference run; the flag becomes part of the saved pipeline config. Claude Code, Codex (including DeepSeek/GLM variants), Gemini CLI, and pi verify the installed executable before extraction and reject mismatches. Install the same exact release separately and use the matching claude_bin, codex_bin, gemini_bin, or pi_bin config key for a non-default installation. Each attempt records its version and executable in traces/<harness>/*.harness.json; successful outputs also include raw_output.harness. Probe failures retain a record with an unknown actual version and the error. For a rerun, install and require the original attempt's recorded version. Omitting the pin records the installed version without requiring one. KYNEX already verifies its fixed Git revision.

Claude Code

Install Claude Code and set an Anthropic API key with access to the selected model:

uv sync --extra anthropic
npm install -g @anthropic-ai/claude-code@X.Y.Z
claude --version
export ANTHROPIC_API_KEY="your-anthropic-api-key"
uv run extract-bench run claude_code_extract_opus_5_5_evidence \
  --dataset pro --max_concurrent 1 --harness_version X.Y.Z

Opus 5.5 requires CLI 2.1.280 or newer. The non-evidence Opus pipelines use --bare, which requires ANTHROPIC_API_KEY or a configured apiKeyHelper; a logged-in Claude session alone is insufficient. The evidence variants do not enable --bare and can also use a logged-in session. Cost comes from the CLI's total_cost_usd.

Codex CLI

Install Codex CLI and set CODEX_API_KEY to an OpenAI API key with access to the selected model:

uv sync --extra openai
npm install -g @openai/codex@X.Y.Z
codex --version
export CODEX_API_KEY="your-openai-api-key"
uv run extract-bench run codex_code_extract_gpt_5_5_low_evidence \
  --dataset pro --max_concurrent 1 --harness_version X.Y.Z

The adapter reads CODEX_API_KEY, rather than the OPENAI_API_KEY used by direct API pipelines. Use a CLI release supporting exec --json --ephemeral --ignore-user-config --ignore-rules. The provider records and verifies the selected version. Evidence variants write page/box citations and use danger-full-access; non-evidence variants use workspace-write. Cost is estimated from reported token usage.

Gemini CLI

Install the selected Gemini CLI version and set a Google API key with access to Argon:

uv sync --extra google
npm install -g @google/gemini-cli@X.Y.Z
gemini --version
export GEMINI_API_KEY="your-google-api-key"
uv run extract-bench run gemini_cli_extract_gemini_4_argon_medium_thinking_vision_evidence \
  --dataset pro --max_concurrent 1 --harness_version X.Y.Z

This pipeline uses medium thinking, an isolated CLI home, local tools and citations.json evidence. The runner enforces a 1,800-second deadline without restarting failed documents. max_cost_usd is a post-hoc flag, not a spending limit.

Requests use Google's native Gemini API. To use a native Gemini-compatible gateway, set GOOGLE_GEMINI_BASE_URL and any required GEMINI_CLI_CUSTOM_HEADERS (name: value, name: value). Credentials remain in the recorder, outside CLI/tool subprocess environments.

Incremental JSONL journals under the output directory's traces/ contain prompts, CLI events, native request/response content, timing, usage and outcomes. Each event is flushed and synced; use a durable output volume to survive machine loss. Costs include helper models, thinking tokens and observed failed-request usage. Missing usage stays unknown: known_cost_usd records the measured subtotal, while total_cost_usd is populated only when accounting is complete. Separate attempts have separate journals; include all applicable attempts when reporting total spend.

pi

Install the pi coding agent and set a Tencent TokenHub API key with access to Hy4:

uv sync --extra local
npm install -g @earendil-works/pi-coding-agent@X.Y.Z
pi --version
export TENCENT_TOKENHUB_API_KEY="your-tokenhub-api-key"
uv run extract-bench run pi_extract_hy4_vision_evidence \
  --dataset pro --max_concurrent 1 --harness_version X.Y.Z

Each request gets its own pi models.json (endpoint, key and prices) outside the agent's working directory, and pi runs without skills, extensions, prompt templates or context files. The pipeline pins reasoning to high, lets the model read rendered pages as images and writes citations.json evidence. It calls TokenHub's OpenAI-compatible endpoint, set by base_url in the pipeline config. Cost is pi's reported cost at TokenHub's list rates; max_cost_usd is a post-hoc flag, not a spending limit.

Cloud API Pipelines

These pipelines use hosted APIs. You only need an API key in your .env file.

Bold pipelines are baselines evaluated in the ParseBench paper. The name used in the paper is shown in parentheses.

LlamaParse

Pipeline Description Env Var
llamaparse_agentic Agentic tier (In paper: LlamaParse Agentic) LLAMA_CLOUD_API_KEY
llamaparse_cost_effective Cost-effective tier (In paper: LlamaParse Cost Effective) LLAMA_CLOUD_API_KEY
llamaparse_agentic_plus Agentic plus tier LLAMA_CLOUD_API_KEY

OpenAI

Pipeline Description Env Var
openai_gpt5_mini_reasoning_medium_parse GPT-5 Mini, medium reasoning, image mode OPENAI_API_KEY
openai_gpt5_mini_reasoning_medium_parse_file GPT-5 Mini, medium reasoning, PDF file mode OPENAI_API_KEY
openai_gpt5_mini_reasoning_minimal_parse GPT-5 Mini, minimal reasoning OPENAI_API_KEY
openai_gpt5_mini_reasoning_minimal_parse_file GPT-5 Mini, minimal reasoning, file mode OPENAI_API_KEY
openai_gpt5_mini_reasoning_medium_parse_with_layout GPT-5 Mini, medium reasoning + layout OPENAI_API_KEY
openai_gpt5_mini_reasoning_medium_parse_with_layout_file GPT-5 Mini, medium reasoning + layout, file (In paper: OpenAI GPT-5 Mini (Reasoning Medium)) OPENAI_API_KEY
openai_gpt5_mini_reasoning_minimal_parse_with_layout GPT-5 Mini, minimal reasoning + layout OPENAI_API_KEY
openai_gpt5_mini_reasoning_minimal_parse_with_layout_file GPT-5 Mini, minimal reasoning + layout, file (In paper: OpenAI GPT-5 Mini (Reasoning Minimal)) OPENAI_API_KEY
openai_gpt_5_4_parse GPT-5.4, image mode OPENAI_API_KEY
openai_gpt_5_4_parse_file GPT-5.4, PDF file mode OPENAI_API_KEY
openai_gpt_5_4_parse_with_layout_file GPT-5.4, parse + layout, file mode (In paper: OpenAI GPT-5.4) OPENAI_API_KEY

Anthropic Claude

Pipeline Description Env Var
anthropic_haiku_parse Claude Haiku 4.5, image mode ANTHROPIC_API_KEY
anthropic_haiku_parse_file Claude Haiku 4.5, PDF file mode ANTHROPIC_API_KEY
anthropic_haiku_parse_with_layout Claude Haiku 4.5, parse + layout ANTHROPIC_API_KEY
anthropic_haiku_parse_with_layout_file Claude Haiku 4.5, parse + layout, file mode (In paper: Anthropic Haiku 4.5 (Disable Thinking)) ANTHROPIC_API_KEY
anthropic_haiku_thinking_parse_with_layout_file Claude Haiku 4.5, extended thinking + layout (In paper: Anthropic Haiku 4.5 (Thinking)) ANTHROPIC_API_KEY
anthropic_opus_4_6_parse Claude Opus 4.6, image mode ANTHROPIC_API_KEY
anthropic_opus_4_6_parse_file Claude Opus 4.6, PDF file mode ANTHROPIC_API_KEY
anthropic_opus_4_6_parse_with_layout_file Claude Opus 4.6, parse + layout, file mode (In paper: Anthropic Opus 4.6) ANTHROPIC_API_KEY
anthropic_opus_4_8_parse_with_layout_file Claude Opus 4.8, parse + layout, file mode (In paper: Anthropic Opus 4.8) ANTHROPIC_API_KEY
anthropic_sonnet_5_parse_with_layout_file Claude Sonnet 5, adaptive thinking + layout, file mode ANTHROPIC_API_KEY
anthropic_fable_5_parse_with_layout_file Claude Fable 5, parse + layout, file mode ANTHROPIC_API_KEY

Google Gemini

Pipeline Description Env Var
google_gemini_3_flash_lite_parse Gemini 3 Flash Lite, image mode GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_lite_parse_file Gemini 3 Flash Lite, file mode GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_thinking_minimal_parse Gemini 3 Flash, minimal thinking GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_thinking_minimal_parse_file Gemini 3 Flash, minimal thinking, file GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_thinking_high_parse Gemini 3 Flash, high thinking GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_thinking_high_parse_file Gemini 3 Flash, high thinking, file GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_thinking_minimal_parse_with_layout Gemini 3 Flash, minimal thinking + layout GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_thinking_high_parse_with_layout Gemini 3 Flash, high thinking + layout GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_thinking_minimal_parse_with_layout_file Gemini 3 Flash, minimal thinking + layout file (In paper: Google Gemini 3 Flash (Thinking Minimal)) GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_thinking_high_parse_with_layout_file Gemini 3 Flash, high thinking + layout file (In paper: Google Gemini 3 Flash (Thinking High)) GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_thinking_minimal_parse_with_layout_agentic_vision Agentic vision, minimal thinking GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_thinking_medium_parse_with_layout_agentic_vision Agentic vision, medium thinking GOOGLE_GEMINI_API_KEY
google_gemini_3_flash_thinking_high_parse_with_layout_agentic_vision Agentic vision, high thinking GOOGLE_GEMINI_API_KEY
google_gemini_3_1_flash_lite_parse Gemini 3.1 Flash Lite GOOGLE_GEMINI_API_KEY
google_gemini_3_1_flash_lite_thinking_high_parse Gemini 3.1 Flash Lite, high thinking GOOGLE_GEMINI_API_KEY
google_gemini_3_1_pro_parse Gemini 3.1 Pro, default thinking GOOGLE_GEMINI_API_KEY
google_gemini_3_1_pro_parse_with_layout_file Gemini 3.1 Pro, parse + layout, file mode (In paper: Google Gemini 3.1 Pro) GOOGLE_GEMINI_API_KEY
google_gemini_3_5_flash_parse_with_layout Gemini 3.5 Flash, default thinking + layout GOOGLE_GEMINI_API_KEY
google_gemini_3_5_flash_no_thinking_parse_with_layout Gemini 3.5 Flash, minimal thinking + layout GOOGLE_GEMINI_API_KEY
google_gemini_3_5_flash_parse_with_layout_file Gemini 3.5 Flash, default thinking + layout, file mode (In paper: Google Gemini 3.5 Flash (Thinking Medium)) GOOGLE_GEMINI_API_KEY
google_gemini_3_5_flash_no_thinking_parse_with_layout_file Gemini 3.5 Flash, minimal thinking + layout, file mode (In paper: Google Gemini 3.5 Flash (Thinking Minimal)) GOOGLE_GEMINI_API_KEY
google_gemini_3_5_flash_lite_parse_with_layout_file Gemini 3.5 Flash Lite, layout + file mode (In paper: Google Gemini 3.5 Flash Lite) GOOGLE_GEMINI_API_KEY
google_gemini_3_6_flash_parse_with_layout_file Gemini 3.6 Flash, default thinking + layout, file mode (In paper: Google Gemini 3.6 Flash (Thinking Medium)) GOOGLE_GEMINI_API_KEY
google_gemini_3_6_flash_no_thinking_parse_with_layout_file Gemini 3.6 Flash, minimal thinking + layout, file mode (In paper: Google Gemini 3.6 Flash (Thinking Minimal)) GOOGLE_GEMINI_API_KEY
google_gemini_3_7_flash_parse_with_layout Gemini 3.7 Flash, default thinking + layout GOOGLE_GEMINI_API_KEY
google_gemini_3_7_flash_no_thinking_parse_with_layout Gemini 3.7 Flash, minimal thinking + layout GOOGLE_GEMINI_API_KEY
google_gemini_3_7_flash_parse_with_layout_file Gemini 3.7 Flash, default thinking + layout, file mode GOOGLE_GEMINI_API_KEY
google_gemini_3_7_flash_no_thinking_parse_with_layout_file Gemini 3.7 Flash, minimal thinking + layout, file mode GOOGLE_GEMINI_API_KEY

Azure Document Intelligence

Pipeline Description Env Vars
azure_di_layout Layout model (In paper: Azure Document Intelligence) AZURE_DOCUMENT_INTELLIGENCE_KEY, AZURE_DOCUMENT_INTELLIGENCE_ENDPOINT
azure_di_read Read model AZURE_DOCUMENT_INTELLIGENCE_KEY, AZURE_DOCUMENT_INTELLIGENCE_ENDPOINT

AWS Textract

Pipeline Description Env Vars
aws_textract Standard Textract (In paper: AWS Textract) AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY
aws_textract_with_forms Textract with forms AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY
aws_textract_text_only Textract text only AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY

Google Document AI

Pipeline Description Env Vars
google_docai Document AI OCR GOOGLE_DOCAI_PROJECT_ID, GOOGLE_DOCAI_PROCESSOR_ID
google_docai_layout Document AI Layout (In paper: Google Cloud Document AI) GOOGLE_DOCAI_PROJECT_ID, GOOGLE_DOCAI_LAYOUT_PROCESSOR_ID

Reducto

Pipeline Description Env Var
reducto Default Reducto (In paper: Reducto) REDUCTO_API_KEY
reducto_agentic Agentic mode (In paper: Reducto (Agentic)) REDUCTO_API_KEY

Pulse

Pipeline Description Env Var
pulse_ultra_2 Pulse Ultra 2 with native markdown output and refinement enabled PULSE_API_KEY

Chunkr

Pipeline Description Env Var
chunkr Default quality CHUNKR_API_KEY
chunkr_high_res High resolution CHUNKR_API_KEY

Datalab (Marker)

Pipeline Description Env Var
datalab_fast Fast mode DATALAB_API_KEY
datalab_balanced Balanced mode DATALAB_API_KEY
datalab_accurate Accurate mode DATALAB_API_KEY

Extend AI

Pipeline Description Env Var
extend_parse Default (In paper: Extend) EXTEND_API_KEY
extend_parse_2 2.0 engine (v2.0.0, GA) EXTEND_API_KEY
extend_parse_light Light engine (v1.0.0) EXTEND_API_KEY
extend_parse_document Document scope EXTEND_API_KEY
extend_parse_section Section scope EXTEND_API_KEY

Landing AI

Pipeline Description Env Var
landingai_parse Default (In paper: LandingAI) LANDING_AI_API_KEY

Unstructured

Pipeline Description Env Var
unstructured_auto Auto strategy UNSTRUCTURED_API_KEY
unstructured_fast Fast strategy UNSTRUCTURED_API_KEY
unstructured_hi_res Hi-res strategy UNSTRUCTURED_API_KEY

OpenInnovation Parser (oi-parser)

Hosted document-parsing API. Sign up at oi-parser.ai to get an API key.

Pipeline Description Env Vars
oi_parser oi-parser hosted /v1/extract API OI_PARSER_API_KEY, OI_PARSER_BASE_URL (optional)

Self-hosted Model Pipelines

These pipelines require you to deploy the model on your own infrastructure (e.g., via vLLM, Modal, etc.) and set the endpoint URL in .env.

Gemma 4

Pipeline Description Env Var
gemma4_26b_vllm Gemma 4 26B-A4B, parse mode GEMMA4_SERVER_URL
gemma4_26b_vllm_with_layout Gemma 4 26B-A4B, layout mode GEMMA4_SERVER_URL
gemma4_e4b_vllm Gemma 4 E4B (dense 8B), parse mode GEMMA4_SERVER_URL
gemma4_e4b_vllm_with_layout Gemma 4 E4B, layout mode GEMMA4_SERVER_URL

Qwen3.5-4B

Pipeline Description Env Var
qwen3_5_4b_vllm_parse Parse mode, markdown (In paper: Qwen 3 VL) QWEN35_SERVER_URL
qwen3_5_4b_vllm_layout Layout mode, JSON with bboxes (In paper: Qwen 3 VL) QWEN35_SERVER_URL

Chandra OCR 2

Pipeline Description Env Var
chandra2_vllm OpenAI-compatible vLLM API CHANDRA2_SERVER_URL
chandra2_sdk Official SDK endpoint CHANDRA2_SERVER_URL

DeepSeek-OCR-2

Pipeline Description Env Var
deepseekocr2_vllm With grounding layout detection DEEPSEEKOCR2_SERVER_URL
deepseekocr2_freeocr Free OCR, no grounding DEEPSEEKOCR2_SERVER_URL

Granite Vision

Pipeline Description Env Var
granite_vision_pipeline PP-DocLayout + per-region Granite Vision GRANITE_VISION_SERVER_URL
granite_vision_4_1_4b Granite Vision 4.1 4B (vLLM, multi-task) VLLM_API_KEY

PaddleOCR-VL

Pipeline Description Env Var
paddleocr_vl_vllm OpenAI-compatible vLLM API PADDLEOCR_SERVER_URL
paddleocr_vl_pipeline Full pipeline (layout + chart routing) PADDLEOCR_SERVER_URL
paddleocr_vl_1_6_vllm PaddleOCR-VL-1.6, OCR prompt PADDLEOCR_SERVER_URL
paddleocr_vl_1_6_vllm_table PaddleOCR-VL-1.6, table recognition prompt PADDLEOCR_SERVER_URL
paddleocr_vl_1_6_pipeline PaddleOCR-VL-1.6, full pipeline (layout + routing) PADDLEOCR_SERVER_URL

dots.ocr

Pipeline Description Env Var
dots_ocr_1_0_parse dots.ocr 1.0 DOTS_OCR_ENDPOINT_URL
dots_ocr_1_5_parse dots.ocr 1.5, layout+text prompt (In paper: Dots OCR 1.5) DOTS_OCR_ENDPOINT_URL

Docling

Pipeline Description Env Vars
docling_parse Docling HTTP endpoint (In paper: Docling) DOCLING_PARSE_ENDPOINT_URL, DOCLING_PARSE_API_KEY (optional)
docling_serve Docling Serve HTTP endpoint DOCLING_SERVE_ENDPOINT_URL, DOCLING_SERVE_API_KEY (optional)

MinerU 2.5

Pipeline Description Env Var
mineru25_vllm MinerU2.5-2509-1.2B vLLM server (two-step layout + recognition) MINERU25_SERVER_URL
mineru2605pro_vllm MinerU2.5-Pro-2605-1.2B vLLM server (adds chart/image analysis) MINERU2605PRO_SERVER_URL

MinerU-Diffusion

Pipeline Description Env Var
mineru_diffusion MinerU-Diffusion-V1-0320-2.5B server (diffusion-decoding OCR, two-stage layout + recognition) MINERU_DIFFUSION_SERVER_URL

Nemotron-Omni

Pipeline Description Env Var
nemotron_omni_30b_vllm_thinking Nemotron-3-Nano-Omni 30B-A3B Reasoning, thinking enabled NEMOTRON_OMNI_SERVER_URL

Surya OCR 2

Pipeline Description Env Var
surya2_sdk Surya OCR 2 SDK server (full-page OCR + layout) SURYA2_SERVER_URL

Local Pipelines (No API key needed)

These run entirely locally with no external dependencies.

Pipeline Description Requirements
pypdf_baseline PyPDF text extraction None
pymupdf_text PyMuPDF text extraction None
pymupdf_html PyMuPDF HTML extraction None
warp_ingest Warp-Ingest local parser warp-ingest[ocr]>=2.0.1 installed
tesseract_eng Tesseract OCR (English) tesseract installed
tesseract_fast Tesseract OCR (fast) tesseract installed
tesseract_high_quality Tesseract OCR (high quality) tesseract installed
infinity_parser2_flash Infinity-Parser2-Flash (vLLM server, JSON layout) infinity_parser2, running vLLM server
infinity_parser2_pro Infinity-Parser2-Pro (vLLM server, JSON layout) infinity_parser2, running vLLM server

Layout Detection Pipelines

Pipeline Description Requirements
docling_layout_heron Docling Heron layout Self-hosted endpoint
docling_layout_heron_101 Docling Heron 1.0.1 Self-hosted endpoint
docling_layout_old Docling legacy layout Self-hosted endpoint
ppdoclayout_plus_l PaddleDetection layout Self-hosted endpoint
qwen3vl_layout Qwen3-VL layout Self-hosted endpoint
surya_layout Surya layout detection surya installed
yolo_doclaynet YOLO DocLayNet Self-hosted endpoint