productionopen access

Platform capabilities

Essential skills, toolkits, and provider adapters

This page makes the platform architecture explicit: what is implemented now, what is pilot-grade, what still needs estimator review, and how provider adapters like GPT or Gemini plug into the workflow.

productiondeterministic-plus-adapters

Strategy

How OpenEstimator is intended to evolve

Document strategy

Deterministic ingestion and document intelligence first, with multimodal providers layered in for harder drawing/spec reasoning.

LLM strategy

OpenEstimator stays provider-agnostic: enable GPT, Gemini, or other adapters per deployment while keeping evidence-first outputs and estimator review in the loop.

Provider strategy

Use the local parsing stack for reproducibility, then route difficult multimodal reasoning tasks to configured providers once keys are present.

Measurement strategy

Counts and evidence-driven quantity support are available today. Geometry-heavy takeoff remains explicitly human-in-the-loop until scale and markup confidence improve.

Workflow

From document ingestion to estimate export

5 stages
1

Workflow stage

Ingest and index documents

active

Upload PDFs, split pages, extract text, and persist document metadata for each project artifact.

Skills

sheet_indexing_skillpage_classification_skill

Outputs

documentspagessheet indexpage thumbnails
2

Workflow stage

Understand drawing structure

active

Classify plans, schedules, specs, details, and addenda while linking page references and title information.

Skills

page_classification_skillschedule_extraction_skillspec_alignment_skill

Outputs

sheet typesschedule candidatesspec links
3

Workflow stage

Extract scope and quantity candidates

active

Generate evidence-backed scope items, map them to CSI, and propose quantities only where support is strong enough.

Skills

scope_detection_skillcsi_mapping_skillquantity_takeoff_skillmeasurement_reasoning_skill

Outputs

estimate itemsevidence referencesmanual review flags
4

Workflow stage

Markup and measurement assist

pilot

Layer future count, length, area, and plan-region markup tools into the review workspace without overstating automation.

Skills

plan_element_extraction_skillcounting_assist_skillscale_calibration_skill

Outputs

markup tracescount regionsmeasurement prompts
5

Workflow stage

Review and export

active

Allow estimators to edit, approve, reject, and export a traceable estimate package with assumptions and audit history.

Skills

estimate_review_skillexport_skill

Outputs

review decisionsexportsaudit trail

Toolkits

Modular estimating building blocks

5 toolkits

Toolkit

Document ingestion toolkit

active

Handles PDF upload, page splitting, thumbnails, OCR/text capture, and structured metadata persistence.

Primary uses

  • Upload multi-file project sets
  • Prepare page-level inputs
  • Normalize storage backends

Tools

PDF uploadpage indexingthumbnail generationtext extractiontable capture

Production-usable for MVP ingestion and fixture-backed demos.

Toolkit

Drawing intelligence toolkit

active

Recognizes sheet types, title information, schedules, spec pages, addenda, and page references.

Primary uses

  • Plan recognition
  • Document structure understanding
  • Drawing/spec linkage

Tools

sheet typingdiscipline detectioncross-reference discoveryschedule identification

Current MVP uses deterministic cues with room to add multimodal fallback providers.

Toolkit

Measurement and takeoff toolkit

pilot

Supports count, length, area, volume, room-rollup, and deduction workflows with explicit review when geometry is uncertain.

Primary uses

  • Quantity support
  • Schedule-driven counts
  • Room and assembly aggregation

Tools

count assistlength assistarea assistvolume assistunit discipline rules

Counts and evidence-driven quantities are present; geometry-based markup is still pilot-grade.

Toolkit

Plan markup toolkit

pilot

Reserved for persistent plan annotations, measurement traces, count regions, and estimator markup overlays.

Primary uses

  • Manual plan markups
  • Measurement traces
  • Linking plan regions to estimate items

Tools

count pinslength tracesarea polygonsopening deduction markersreview callouts

UI scaffolding is surfaced, but persistent geometry and region storage are still a next milestone.

Toolkit

Review and export toolkit

active

Packages the estimate into reviewable tables, exports, audit trails, and benchmark-friendly outputs.

Primary uses

  • Estimate review
  • Audit and compliance traceability
  • Workbook/export generation

Tools

review tableaudit trailcsv/json exportsassumption reportambiguity list

Operational in the MVP today and ready to expand into comments and collaboration.

Providers

External and local model pathways

5 providers

Model provider

Local document intelligence stack

active

This is the default runtime path for ingestion, page previews, and baseline document parsing.

pdf-ocr-cvConfiguredLocal runtime

Best at

  • PDF splitting and page indexing
  • Text extraction and table capture
  • Deterministic fallback when no LLM provider is configured

Default models

PyMuPDFpdfplumberPillow

Model provider

OpenAI

adapter-ready

Recommended for document-grounded reasoning and estimator-facing explanations once configured.

multimodal-llmNot configuredAPI key

Best at

  • Multimodal drawing and specification reasoning
  • Evidence-grounded ambiguity explanation
  • Agent orchestration for scope extraction

Default models

gpt-5gpt-4.1

Environment variable

OPENESTIMATOR_OPENAI_API_KEY

Model provider

Google Gemini

adapter-ready

Positioned as an interchangeable multimodal provider for page reasoning and benchmark workflows.

multimodal-llmNot configuredAPI key

Best at

  • Large drawing-set summarization
  • Fast sheet-level triage
  • Multimodal page and schedule understanding

Default models

gemini-2.5-progemini-2.5-flash

Environment variable

OPENESTIMATOR_GEMINI_API_KEY

Model provider

Anthropic

adapter-ready

Useful for disciplined trade-scope reasoning once credentials are provided.

llmNot configuredAPI key

Best at

  • Long-context specification reasoning
  • Assumption and exclusion drafting
  • Estimator-facing explanation quality

Default models

claude-sonnetclaude-opus

Environment variable

OPENESTIMATOR_ANTHROPIC_API_KEY

Model provider

Vision geometry toolkit

pilot

This is the modular slot for computer-vision-assisted count, measurement, and markup extraction.

cv-geometryConfiguredLocal runtime

Best at

  • Plan element extraction scaffolding
  • Future count and geometry assist modules
  • Legend and symbol alignment candidates

Default models

linework calibrationsymbol spottingregion proposal

Skills

Estimating agent skill registry

6 skills

document-intelligence

Sheet indexing

active

Builds a cross-sheet index from title blocks, sheet labels, and drawing index pages.

deterministic2 toolkit links1 provider paths

Inputs

page textsheet labelsdrawing index pages

Outputs

sheet indexcross-referencessheet number normalization

Review triggers

  • missing sheet numbers
  • duplicate sheet IDs

document-intelligence

Page classification

active

Classifies pages as plans, schedules, specs, addenda, details, or legends with discipline hints.

deterministic1 toolkit links3 provider paths

Inputs

page textsheet identifiersdocument metadata

Outputs

sheet typedisciplineclassification evidence

Review triggers

  • mixed-use sheets
  • unclear discipline

estimating-reasoning

Scope detection

active

Converts document cues into candidate estimate items while preserving direct evidence vs inference.

evidence-assisted1 toolkit links4 provider paths

Inputs

classified pagesschedule evidencespec linksnotes

Outputs

detected scope itemstraceable evidence snippets

Review triggers

  • inferred scope
  • trade overlap
  • missing quantity

measurement

Quantity takeoff

active

Assigns counts, lengths, areas, or volumes only when evidence is explicit enough to support a safe proposal.

evidence-assisted1 toolkit links3 provider paths

Inputs

scope candidatesschedule rowstextual quantities

Outputs

quantity candidatesunit formulasmanual review flags

Review triggers

  • missing scale
  • opening deductions needed
  • partial quantity evidence

vision-recognition

Plan element extraction

pilot

Finds candidate rooms, wall runs, openings, symbols, and repeated assemblies from 2D drawing regions.

multimodal-agent2 toolkit links3 provider paths

Inputs

page imageslegendstitle blockssymbol cues

Outputs

region candidatesplan element proposalsmarkup seeds

Review triggers

  • symbol ambiguity
  • legend missing
  • poor raster quality

measurement

Counting assist

pilot

Supports repetitive counts for doors, fixtures, devices, and keyed symbols with estimator confirmation.

human-in-loop2 toolkit links3 provider paths

Inputs

schedule rowssymbol detectionsmanual markups

Outputs

count regionscount formulasreview checklist

Review triggers

  • occluded symbols
  • legend mismatch
  • repeated details