Warehouse Label Matching From Existing Text Inference

# Freelancer Requirement: Warehouse Label Matching From Existing Text Inference ## Project Context We already have these components working in production: - QR code detection - Barcode detection - Text detection / OCR inference We do **not** need a new OCR model or a new QR/barcode detector. The required work is to build/improve a **warehouse label matching module** that takes existing OCR text results and matches them to warehouse label records in our inventory data. ## Goal Given OCR text detections from images, match text candidates to the correct warehouse label with high precision and practical recall, and return ranked matches with confidence and explainability. ## In Scope - Use existing text inference outputs (detected text, confidence, bounding boxes) as input. - Normalize noisy OCR strings (e.g., O/0, I/1, B/8, S/5 confusion). - Match OCR text candidates to warehouse labels in inventory. - Support exact and fuzzy/partial matching. - Return top-N ranked candidates with confidence score and match reason. - Handle ambiguity (multiple similar labels) and produce confidence flags. - Provide clear thresholds/configs for production tuning. - Integrate into the current Python pipeline with minimal disruption. ## Out of Scope - Building a new OCR/text detection model. - Building or replacing QR/barcode pipelines. - UI/frontend work. - Full data platform redesign. ## Existing Inputs and Data Expected OCR input per text detection (already available): - `text`: OCR text string - `confidence`: OCR confidence - `bbox`: bounding box coordinates Data we will provide to the freelancer for implementation and validation: - A representative set of real inference results (JSON/text outputs) - Original images from production-like runs - Relevant image crops for difficult label regions (including split/wavy labels) - A copy/snapshot of actual warehouse inventory records to match against (with required label fields) Expected inventory/label fields (examples): - `LicensePlateNumber` / `license_plate_number` - `ItemNumber` - `Location` - Other metadata used for tie-breaking and audit ## Functional Requirements 1. Candidate Extraction - Parse OCR text detections and generate normalized label candidates. - Support fragmented or partial label reads. - Reconstruct labels from split OCR fragments when text is broken into multiple detections (for example wavy/curved labels or damaged print). - Use geometric and text heuristics to group nearby fragments (same row/line, overlap/proximity, reading order) and generate merged candidates. - Try multiple join strategies (left-to-right concatenation, separator removal, OCR-corrected merge) before matching. 2. Matching Engine - Perform exact pattern match where possible. - Perform fuzzy/partial matching for incomplete text. - Match both single-fragment and merged-fragment candidates. - Score matches with a composite method (string similarity, edit distance, positional weighting, sequence proximity when applicable). - Return top 3-10 matches per candidate with scores. 3. Confidence + Ambiguity Handling - Assign confidence tiers (for example: HIGH, MEDIUM, LOW, AMBIGUOUS). - Detect ambiguous prefix/similar-label cases and downgrade confidence. - If sequential context/history is available, use it as an optional boost. 4. Output Contract For each accepted text-derived match, return: - matched label - source type (exact/fuzzy) - confidence score - confidence flag - reason/explanation (why this was selected) - linked inventory metadata - candidate construction metadata (single OCR text vs merged from fragments, and which fragment indices were joined) 5. Fallback Behavior - If no reliable match exists above threshold, return `NO_MATCH` with debug info. - Never overwrite higher-priority matches from QR/barcode flows. ## Non-Functional Requirements - Deterministic results for same input. - Configurable thresholds via environment/config values. - Efficient runtime suitable for Lambda-style execution. - Logging for observability and debugging. - Unit-testable design with pure matching functions. ## Deliverables 1. Production-ready Python module - Clean, documented code for normalization + matching + ranking. 2. Integration changes - Hooked into current inference pipeline where OCR text is already processed. - Backward-compatible output structure where feasible. 3. Test suite - Unit tests for exact, fuzzy, ambiguous, and no-match cases. - Fixture-based tests using representative warehouse labels and OCR noise patterns. - Include tests where true labels are split across 2-3 OCR detections and must be joined to match correctly. 4. Evaluation report - Precision/recall-style metrics on provided sample set. - Error analysis: false positives vs false negatives. - Recommended threshold defaults and rationale. - Validation must be performed on the provided real inference results, images, crops, and warehouse inventory snapshot (not only synthetic examples). 5. Handover notes - Configuration guide - How to run tests - How to tune thresholds safely ## Acceptance Criteria - Correctly matches known sample labels from OCR outputs at agreed target accuracy. - Ambiguous matches are flagged instead of silently misclassified. - No regressions to existing QR/barcode pipelines. - Code passes tests and is deployable in current repo structure. - Split-text labels (broken OCR regions) are correctly reconstructed and matched in agreed sample scenarios. ## Suggested Technical Approach - Text normalization layer for OCR artifacts. - Fragment grouping + reconstruction layer before matching (spatial clustering + reading-order join). - Pattern-aware matcher for known label formats. - Composite scoring combining: - normalized string similarity - weighted edit distance - suffix/prefix sensitivity - optional sequential proximity (if labels are sequential) - Rank + threshold + confidence policy layer. ## Timeline and Commercial Terms - Engagement type: **Fixed price only** (no hourly proposals). - Priority: **ASAP delivery**. - Expected start: immediately after award. - Freelancer must propose a fast delivery plan with concrete dates. - Preferred target: first working version in 2-4 days, final handover in 5-7 days (unless a justified alternative is approved). ## Skills Required - Strong Python experience - OCR post-processing and noisy text matching - Similarity algorithms (Levenshtein / weighted scoring) - Practical MLOps mindset for production inference pipelines - Experience with AWS Lambda/S3-style workflows is a plus ## Proposal Instructions for Freelancer Please include in your proposal: - Similar projects where you matched noisy OCR text to structured inventory labels - Proposed scoring/matching strategy - Testing strategy and an **ASAP** timeline with dates - **Fixed-price quote** with clear scope assumptions (no hourly quote) - Any assumptions or risks you see in this scope

Project ID: 40511239