Release Notes

Track new features, improvements, and fixes for the APSIS Document Analysis API.

---

v1.3.0 February 2026

• Drawing Classification - Automatic work package, drawing type, and summary extraction for drawings

API Changes (Action Required)

These changes may require updates to your integration:

New Fields on Drawing info

Four new nullable fields are now returned in the info object for drawings:

Field	Type	Description
work_package_code	string or null	Work package code (e.g., "2000") from a predefined list of construction categories
work_package_description	string or null	Human-readable description (e.g., "Superstructure - Frame")
drawing_type	string or null	One of: PLAN, ELEVATION, CONSTRUCTION DETAILS, SPECIFICATION
summary	string or null	Brief AI-generated summary of the drawing contents

These fields are only populated for documents classified as drawings (not specifications). They will be null if classification was skipped or failed.

---

How It Works

Classification runs automatically as part of the Extract Data pipeline (no additional API calls needed). For each drawing:

1. Azure Document Intelligence extracts text via OCR
2. LLM classifies the drawing against a predefined work package list and determines the drawing type
3. Results are merged into the existing info response

Adds roughly 10 seconds to extraction time per drawing.

---

v1.2.0 January 2026

• Token Usage Tracking - Per-request token counts and cost reporting on all API operations
• Model Selection for Drawing Compare - Choose between OpenAI and Anthropic models
• Image-Based Specification Extraction - Extract Data now uses OCR image analysis for specifications, removed Docling approach to improve accuracy
• Increased PDF Page Limit - Extract Data/Doc Search now supports PDFs up to 2,000 pages (was 50) due to client request of having a requirement to use Doc Search on large documents

API Changes (Action Required)

These changes may require updates to your integration:

1. Specification Author Field Type Change

The author field in spec_info is now an array (matching the drawings info.author change from v1.1.0).

Before (v1.1.0)	After (v1.2.0)
"author": "Jane Doe"	"author": ["Jane Doe"]

2. New Request Parameters

Endpoint	Parameter	Type	Default	Description
POST /api/drawings/agent-comparisons	model	string	gpt-5.1	LLM model for change analysis

Supported models include OpenAI (e.g. gpt-5.1, gpt-5.2) and Anthropic with litellm/anthropic/ prefix (e.g. litellm/anthropic/claude-3-5-sonnet-20241022).

---

New Features

Token Usage Tracking

Every API response now includes a token_usage field with a full breakdown of AI model usage and estimated cost. This is logged to the API and returned in responses for all features:

• Extract Data: token_usage on the task object (GET /api/drawings/tasks/{id})
• PDF Diff: token_usage on the diff object (GET /api/drawings/pdf-diffs/{id})
• Drawing Compare: token_usage on the comparison object (GET /api/drawings/agent-comparisons/{id})
• Chunk & Embed: token_usage on the embedding task (GET /api/drawings/embedding-tasks/{id})
• Document Search (RAG): token_usage returned inline in the response (POST /api/drawings/rag)

Token usage is tracked even on failed tasks — any tokens consumed before the failure are recorded.

Example token_usage field:

{
  "token_usage": {
    "models": {
      "gpt-4o": { "input_tokens": 25000, "output_tokens": 4000, "total_tokens": 29000 },
      "gpt-4o-mini": { "input_tokens": 1200, "output_tokens": 300, "total_tokens": 1500 },
      "text-embedding-3-small": { "input_tokens": 12000, "output_tokens": 0, "total_tokens": 12000 }
    },
    "total_cost_usd": 0.103465
  }
}

See Token Usage Tracking in the API docs for full details on viewing and collecting usage data.

Model Selection for Drawing Compare

You can now select which LLM model to use when creating a drawing comparison. This supports both OpenAI and Anthropic models via LiteLLM.

curl -X POST ".../api/drawings/agent-comparisons" \
  -d '{"prev_file_id": 123, "new_file_id": 125, "model": "gpt-5.2"}'

Image-Based Specification Extraction

Specifications are now extracted using OCR image analysis (matching the approach already used for drawings), replacing the previous Docling-based extraction. This results in more accurate author detection, document titles, and revision information from specification PDFs.

---

Improvements

Performance

• PDF page limit increased from 50 to 2,000 pages per document
• Chunk embedding processes in batches for better memory handling on large documents

Drawing Compare

• Faster processing with optimized image alignment
• Better accuracy with improved change detection and reduced false positives
• Support for both OpenAI and Anthropic models

Accuracy

• OCR image-based extraction for specifications replaces Docling, improving metadata quality
• Compare accuracy improvements with updated prompt engineering

---

v1.1.0 January 2026

• Multiple Authors - Now detects all company names on drawings, not just the first
• Confidence Indicators - Know when extractions are high or low confidence
• Faster Processing - Parallel batch processing with reduced timeouts
• LangSmith Tracing - Full visibility into every LLM call with cost dashboards

API Changes (Action Required)

These changes may require updates to your integration:

1. Author Field Type Change

The author field in drawing info is now an array to support multiple authors. (The spec_info.author field was also changed to an array in v1.2.0.)

Before (v1.0.0)	After (v1.1.0)
"author": "Smith & Associates"	"author": ["Smith & Associates", "Jones Engineering"]

2. New Fields Added

Field	Type	Description
extraction_method	string	yolo_crop (high confidence), heuristic_crop (low confidence), or null

3. Date Format Standardised

Dates are now returned in DD-MM-YYYY format (UK standard).

Before (v1.0.0)	After (v1.1.0)
"2024-01-15" or "January 15, 2024"	"15-01-2024"

---

New Features

Multiple Authors Detection

The system now detects and returns all company names and authors found on a drawing, not just the first one. Useful for drawings involving multiple firms (e.g., architect + structural engineer).

Extraction Confidence Indicator

Each extracted drawing includes a confidence level:

Confidence	Meaning
High (yolo_crop)	Object detection model found and cropped the title block area
Low (heuristic_crop)	Title block not detected; used fallback (bottom 50% of page)

Low confidence extractions may be less accurate but still provide useful data in most cases. Standard formats with title blocks in the bottom right typically extract well.

Smarter Filename Recognition

The system uses PDF filenames as hints for author detection. Current mappings:

Filename Pattern	Author
271_*	Convery Prenty Architects
649*	Designme
22007_* or 21014_*	Apsis
7871_*	Grossart Associates
3077-CDP-*	Clyde Design Partnership
1477-ABC-*	Anderson Bell + Christie

We can add more mappings as needed - just let us know.

Automatic Search Indexing

Documents are automatically indexed for Document Search after extraction - no additional steps required.

---

Improvements

Performance

• Documents processed in parallel batches, significantly reducing wait times
• Timeout reduced from 2 minutes to 1 minute per document
• Results returned immediately while embedding (indexing) happens in background

Accuracy

• Upgraded to GPT-5.2 model for reading title block information
• Improved padding around title blocks to avoid cutting off edge fields
• Better classification to distinguish Drawings from Specifications
• Author names and drawing titles formatted in Title Case
• Dates standardised to DD-MM-YYYY format

Reliability

• Long text fields handled gracefully instead of causing errors
• Better recovery from temporary failures

---

Bug Fixes

• Fixed issue where very long drawing titles could cause extraction to fail
• Improved handling of unusual drawing formats

---

Cost & Transparency

LangSmith Tracing

Every API call is also tracked via LangSmith, enabling:

• Full trace visibility for every LLM call
• Daily/weekly/monthly usage dashboards
• Cost analysis broken down by document or time period

Cost Example

Processing 6 construction drawings:

File
0195 BW-SL-010—LANDSCAPE LAYOUT.pdf
0195 BW-SL-011-A-FINISHES & BOUNDARY TREATMENT.pdf
0195 BW-SL-012-A-PLOT WORKS PLOTS 1-7.pdf
0195 BW-SL-013—PLOT WORKS PLOTS 8-17.pdf
0195 BW-SL-014—PLOT WORKS PLOTS 18-34.pdf
0195 Newburgh Issue Sheet-1.0-SB Architects -DET.pdf

Metric	Value
Files processed	6 drawings
Total tokens	30,862
Total cost	$0.08
Cost per drawing	~$0.01

Cost Optimisations

• Classification images compressed, reducing API costs by ~95%
• Large drawings automatically resized before processing

---

Technical Notes

Fine-Tuning Research

We trained a custom APSIS model for title block extraction. While accurate, it was significantly slower than the standard model. After weighing trade-offs, we chose GPT-5.2 for its similar quality and faster processing times.

Current Infrastructure

Model Hosting: The system uses a Tier 4 GPT-5.2 model hosted on the Agency AI OpenAI account, providing high rate limits (2M tokens/min) with pay-as-you-go billing.

Future Options To Consider:

Option	Pros	Cons
APSIS OpenAI Account	Direct billing to APSIS, same API	Requires $250 spend to reach Tier 4 rate limits
APSIS Azure OpenAI	UK data residency, Azure AD integration, enterprise SLA	Lower default rate limits (must request increases from Microsoft)

Key Trade-offs:

• OpenAI Direct: Higher rate limits automatically, latest models first, but data processed in US
• Azure OpenAI: Better for UK/EU compliance and enterprise integration, but slower rate limit approvals

Rate Limiting: Currently no per-user limits are enforced. Options to consider include per-user token quotas, concurrent request limits, or hosting multiple model instances with round-robin distribution. Need to have a discussion with Alex/Paul on this asap.

---

v1.0.0 November 2024

• Initial release of the APSIS Document Analysis API

Features:

• Extract Data - Automatically extract metadata from drawings and specifications
• Detect Revision - Track document versions and predict next revision numbers
• Compare Specifications - Visual diff and AI analysis of specification changes
• Document Search - Semantic search using natural language
• Drawing Compare - Visual comparison of construction drawings