Category id 55d08ff6d3b4207b (Scout category id). 11 alternatives, shown in Scout's order.
Scout's category view
Scout's verdict: evaluate. llm_aided_ocr is dicklesworthstone's most-starred tool (2,866). LLM-enhanced OCR for the Pixel pipeline.
Org setup at the time, as Scout recorded it: Tesseract (basic) + manual PDF extraction
Scout's judgement, from alternatives.json (2026-11-18). It is not a gaents rating.
Our own measured runs
No head-to-head measured in this category yet.
Alternatives
llm_aided_ocr
open-sourcenot runnable on gaents yet
llm_aided_ocr improves OCR accuracy by using LLMs to post-process Tesseract output. Feeds raw OCR text to an LLM with visual context to correct errors, disambiguate ambiguous characters, and infer layout structure.
Pipeline: Tesseract OCR → LLM correction → cleaned output. Works best on documents with good image quality. Supports Claude, GPT-4, and other LLMs. Configurable confidence thresholds and correction... (Scout's description)
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
NOT INSTALLED. LLM-enhanced OCR. High relevance for document pipelines. ⚠ LICENCE RECORD CORRECTED 2026-08-20: our DB said plain 'MIT'. The actual LICENSE on this repo is "MIT License (with OpenAI/Anthropic Rider)". The rider defines Restricted Parties as OpenAI L.L.C., ANTHROPIC PBC, their affiliates, and anyone 'acting directly or indirectly on behalf of, for the benefit of, or under the direction of' them — and grants NO RIGHTS to any Restricted Party. MY READ (non-lawyer, routed to Lex): we are almost certainly NOT a Restricted Party — we are an Anthropic CUSTOMER, and buying a service is not acting on the vendor's behalf or under its direction. But 'for the benefit of' is broad, we build agents on Claude, and I am not qualified to close it. Recording the true licence text rather than the convenient summary. · 2026-09-14 LEX RULING (licence cleared): MCL is a CUSTOMER of Anthropic, NOT a Restricted Party under the OpenAI/Anthropic rider. Buyer-seller is not agency, not direction, and not 'for the benefit of' on any construction surviving commercial common sense (Rainy Sky v Kookmin [2011] UKSC 50; contra proferentem per Investors Compensation Scheme [1998]; noscitur a sociis). Licence VALID for our use. Confidence HIGH. Full ruling: ~/Projects/lex/outbox/2026-09-14-scout-anthropic-rider-ruling.md. CARVE-OUTS (not live, recorded so a future audit need not rebuild them): (1) a contractor/reseller/JV engagement WITH Anthropic or OpenAI would engage the rider for that engagement's workload; (2) redistribution TO a Restricted Party as end recipient is void. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class source-available because LICENSE is 'MIT License (with OpenAI/Anthropic Rider)': no rights for OpenAI/Anthropic and parties acting for them (agent-chose class; stored type left as is, Scout to rule). self_hostable = local Python install (README). api_or_webhooks left out (script, no documented API). data_export n/a (stateless).]
Why this link: the API's job list (GET /v1/jobs, read 2026-11-19T06:07:32Z) names this item as the job's source (proven_item_id).
Tesseract is Google's open-source OCR engine supporting 100+ languages. Combines classical OCR (feature extraction) with neural networks for recognition. Outputs text, PDF, TSV, or hOCR with character confidence scores.
Works well on scanned documents with clear text. Less effective on handwriting, complex layouts, or low-quality images. Command-line tool with library bindings (C, Python, Node.js, etc.).
Stable... (Scout's description)
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
Industry standard OCR. llm_aided_ocr wraps Tesseract. Used as dependency. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from LICENSE (Apache-2.0); self_hostable and api_or_webhooks from README (library + CLI); data_export n/a (library).]
PaddleOCR
open-sourcenot runnable on gaents yet
PaddleOCR is Baidu's OCR system supporting 100+ languages with better multi-language and complex layout support than Tesseract. Uses deep learning (PP-OCR architecture) for text detection and recognition.
Lightweight (~500MB including models). Fast on CPU; supports GPU acceleration. Good performance on tables, forms, and handwriting. Includes text detection and layout analysis (character boxes, line... (Scout's description)
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Superior multi-language and complex layouts. Lightweight. Preprocessing for llm_aided_ocr. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from LICENSE (Apache-2.0); self_hostable from README (open-sourced deployment, Docker images); api_or_webhooks from README (SDKs); data_export n/a (OCR library).]
LangExtract
open-sourcenot runnable on gaents yet
LangExtract is a Python library from Google that wraps LLM calls (Gemini, OpenAI, Anthropic, or any OpenAI-compatible local endpoint) to perform schema-guided structured entity extraction from unstructured text. You define a Pydantic or JSON schema describing the entities and relationships you want, pass in raw text, and it handles prompt construction, chunking for long documents, retry logic on malformed output,... (Scout's description)
~£50 over 2 years for our usage. Software is free (Apache-2.0). No hardware requirements beyond what we already run. Real cost is LLM inference (which we pay regardless of tool choice) plus integration time: ~6 hours one-off, ~1 hour/month maintenance = ~30 hours over 2 years. If we use it to replace ad-hoc JSON-extraction prompts across 90 agents, it likely saves more engineering time than it costs.
Adopt now for document-ingestion and knowledge-graph workflows where span attribution matters; keep Instructor or raw JSON mode for simple cases. Integration cost is low, the library is well-engineered, and 35.5k stars signals it will be maintained.
Limits for us
Python-only — our Rust or TypeScript agents cannot use it directly and would need a service wrapper. The chunking-with-context approach adds token overhead (~15-25% vs naive chunking). Span attribution only works when the LLM preserves exact source substrings; models that paraphrase will return empty spans. For our 90-agent fleet, the main limitation is that most agents do not need heavyweight extraction. Best fit is specific workflows (document ingestion, knowledge graph building, compliance review).
Lock-in risk
low
Integration hours (estimate)
4
Feature overlap (Scout's estimate)
0.4
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
LLM-based structured entity extraction from text, NOT a PDF parser. Precise source grounding (char offsets). Interactive HTML viz. Works with Gemini, OpenAI, withheld. Complements Docling/Marker (they convert docs to text, LangExtract extracts structured data from text). 128KB pure Python package. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from LICENSE (Apache-2.0); self_hostable and api_or_webhooks from README (Python library); data_export n/a (stateless library).]
franken_ocr
source-availablenot runnable on gaents yet
Pure-Rust, CPU-only OCR engine running a DeepSeek-OCR-derived 3B MoE VLM (Baidu Unlimited-OCR). Five-model zoo, custom int8 kernels, no ML framework, no Python, no GPU. (Scout's description)
CPU-only + no-GPU + no-Python-dependency is genuinely attractive for our VPS stack (no GPU boxes). NOASSERTION license needs clearing before adoption. Relevant to Pixel's OCR needs. Evaluate vs incumbent PaddleOCR/Tesseract on accuracy + throughput on our CPU boxes.
Lock-in risk
low
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
⚠ LICENCE RECORD CORRECTED 2026-08-20: our DB said plain 'MIT'. The actual LICENSE on this repo is "MIT License (with OpenAI/Anthropic Rider)". The rider defines Restricted Parties as OpenAI L.L.C., ANTHROPIC PBC, their affiliates, and anyone 'acting directly or indirectly on behalf of, for the benefit of, or under the direction of' them — and grants NO RIGHTS to any Restricted Party. MY READ (non-lawyer, routed to Lex): we are almost certainly NOT a Restricted Party — we are an Anthropic CUSTOMER, and buying a service is not acting on the vendor's behalf or under its direction. But 'for the benefit of' is broad, we build agents on Claude, and I am not qualified to close it. Recording the true licence text rather than the convenient summary. · 2026-09-14 LEX RULING (licence cleared): MCL is a CUSTOMER of Anthropic, NOT a Restricted Party under the OpenAI/Anthropic rider. Buyer-seller is not agency, not direction, and not 'for the benefit of' on any construction surviving commercial common sense (Rainy Sky v Kookmin [2011] UKSC 50; contra proferentem per Investors Compensation Scheme [1998]; noscitur a sociis). Licence VALID for our use. Confidence HIGH. Full ruling: ~/Projects/lex/outbox/2026-09-14-scout-anthropic-rider-ruling.md. CARVE-OUTS (not live, recorded so a future audit need not rebuild them): (1) a contractor/reseller/JV engagement WITH Anthropic or OpenAI would engage the rider for that engagement's workload; (2) redistribution TO a Restricted Party as end recipient is void. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: CORRECTION: type open-source -> source-available and license text clarified: LICENSE is MIT plus a rider that grants no rights to named parties (a use restriction), so licence_class source-available. self_hostable from install script; api_or_webhooks from embeddable Rust API; data_export n/a (OCR tool).]
pdf-inspector (Firecrawl)
open-sourcenot runnable on gaents yet
Fast Rust library for PDF classification (scanned vs text-based vs image-based vs mixed) and text extraction with markdown conversion. Does NOT perform OCR — classifies PDFs and routes scanned pages to external OCR services while handling text-based PDFs locally. Built by Firecrawl to skip expensive OCR for the ~54% of PDFs that are already text-based. Extracts positioned text with font info, detects tables via... (Scout's description)
Best-in-class fast PDF classifier/router. Complements OCR tools (marker, Docling, MinerU) rather than replacing them. Standalone from Firecrawl ecosystem. Very relevant for Pixel's iXBRL/OCR pipeline as a fast pre-filter.
Limits for us
Does NOT do OCR — classifies and routes only. RTL text (Hebrew/Arabic) has known issues (spaces between characters, visual vs logical order — active PR #334). Regression on main for form-field PDFs (#342, fixed in 0.2.6 release). No image/figure content extraction. No formula recognition.
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
[2026-10-20T01:33Z cloud refresh: type None -> open-source (license text: MIT; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from LICENSE (MIT); self_hostable (library) and api_or_webhooks from README. data_export n/a (stateless library).]
AI-reads-books-page-by-page (echohive42)
open-sourcenot runnable on gaents yet
Performs page-by-page analysis of PDF books to extract knowledge points and generate progressive summaries. Solo dev repo. (Scout's description)
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Category mismatch signal: this is really 'reading' not 'analytics'. Kept as skip because the underlying approach is dominated by better tools. [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: last commit 2026-06-27; licence unchanged (MIT).]
Docling
open-sourcenot runnable on gaents yet
IBM DS4SD's PDF-and-doc extraction toolkit. Layout understanding, table structure recognition, reading-order preservation, structured output. The current state-of-the-art OSS PDF-to-structured-doc converter — supplants pdfplumber-only pipelines for anything with tables or multi-column layout. (Scout's description)
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
Missed in S30 + S32 DB sweeps despite being 65K+ stars — real coverage failure caught by S35 ultrathink Fork C. IBM commercial-grade eng but MIT-licensed so no rider concerns. Recommended primary layer 2 in pdf-text v2 rewrite. [2026-10-19T22:08Z cloud refresh: Q4-2026 review: repo moved from DS4SD to docling-project (PyPI project_urls). Last commit 2026-10-19; latest PyPI 2.140.0. MIT unchanged.] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from LICENSE (MIT); self_hostable (pip library) and api_or_webhooks from README; data_export n/a (library).]
weights under modified OpenRAIL-M (code Apache-2.0)
Scout, gaents-proven-2026-11-18.md A22026-11-18Scout review date
Scout's judgement
Verdict
adopt
Lock-in risk
low
Integration hours (estimate)
4
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
Missed in DB along with Docling. Same VikParuchuri also authored pdf-inspector (14.5K stars, evaluate). Marker is the higher-star sibling and paper-specific. [2026-10-19T22:08Z cloud refresh: Q4-2026 review: licence field now records the weights term. Last commit 2026-10-19. Verdict unchanged.] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class source-available (agent-chose): code is Apache-2.0 but the model weights it needs use a modified OpenRAIL-M licence with a $5M funding/revenue limit. self_hostable from README pip install; api_or_webhooks from README (Python converter classes, CLI); data_export n/a (stateless converter).]
Why this link: the API's job list (GET /v1/jobs, read 2026-11-19T06:07:32Z) names this item as the job's source (proven_item_id).
Microsoft's format-native converter: almost any file to clean Markdown. PDF, Word, PowerPoint, Excel, images, audio, HTML and ZIP. No ML in the default path -- it reads each format with a native parser, which is why it returns in seconds where layout-ML tools take minutes. Optional LLM client for image descriptions and an optional Azure Document Intelligence backend, both off by default. (Scout's description)
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout tested it (full, 2026-09-14): Verify THE claim that justified ranking it ahead of Docling as tier 1: that it handles PowerPoint / Excel / ZIP, which Docling, Marker and the retired pdf-text button all cannot.
Scout's notes
HEADLINE CHECK: the tweet said '100,000+ stars' -- actual 183,976, UNDERSTATED by 84K. 'Official Microsoft open-source tool, free, commercially usable' is TRUE (MIT). The '200-page PDF, shockingly clean' line is one anecdote and the tweet ends in a funnel to the author's paid workshop: the tool is real, the framing is lead-gen. | WHY THIS IS A REAL GAP: S36 added Docling (adopt) and Marker (adopt) to this shelf after Adam's pdf-text pushback. MarkItDown is the third member of that family and was missed. It is not a Docling competitor -- it is the CHEAP TIER of the same stack. Approach: format-native parsers vs layout ML. Speed: seconds vs minutes. Formats: adds PowerPoint, Excel, images, audio and ZIP, none of which our current PDF path can touch at all. Quality on messy scanned PDF: weak, which is exactly where Docling earns its cost. | THE CORRECT ARCHITECTURE IS A CASCADE, and MarkItDown is stage one: MarkItDown for everything -> escalate to Docling only when output is empty or garbled -> Marker for the residue. Running Docling on a .docx is paying ML prices to read a zip of XML. | AFFECTS WORK ALREADY SCHEDULED: the pdf-text button retirement is mid-flight. The replacement must route MarkItDown-first, not Docling-first, or we replace one over-engineered button with another. | TUNABLES: install extra / `markitdown[all]` / USE THE BASE INSTALL / [all] pulls Azure Document Intelligence and audio-transcription dependencies we do not want on the document path. --llm-client for image descriptions / off / LEAVE OFF / it is a per-image token bill and it is Docling's job. Azure DI endpoint / unset / LEAVE UNSET / setting it sends our documents to Azure. --use-plugins / off / LEAVE OFF / third-party plugins on the document path is a supply-chain surface. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from LICENSE (MIT); self_hostable and api_or_webhooks from README (Python library and CLI); data_export n/a (stateless library).]
lift (Datalab)
open-sourcenot runnable on gaents yet
See research/tweet-triage-2026-10-14-scout.md (section for lift (Datalab)). (Scout's description)
One triage pass; re-verify stars, licence and price before acting.
Limits for us
See research/tweet-triage-2026-10-14-scout.md (section for lift (Datalab)).
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Merged from research/tweet-triage-2026-10-14-scout.md by merge_triage_table.py. [2026-10-20T01:33Z cloud refresh: type None -> open-source (license text: Apache; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class: mixed row; code LICENSE is Apache-2.0 but README says commercial self-hosting of the model needs a licence (weights under modified OpenRAIL-M), so class of the core product (model) = source-available; self_hostable from pip install; api_or_webhooks true (Python package); data_export n/a.]