Indic document intelligence benchmark
A reproducible benchmark for comparing OCR, information retrieval, and question-answering systems on multilingual Indian administrative documents.
Project definition
Problem statement
Indian administrative documents contain multiple languages, varied scripts, scanned pages, tables, stamps, and inconsistent layouts. A model that works on clean English text may perform poorly on these documents.
The engineering problem is to build a controlled benchmark that separates OCR, retrieval, and question-answering performance so that systems can be compared using the same documents, questions, ground truth, and metrics.
Project objectives
- Prepare a versioned corpus of permitted multilingual administrative documents.
- Create ground truth for page text, document fields, retrieval passages, and question answers.
- Compare selected OCR, retrieval, and question-answering approaches under identical conditions.
- Measure results by language, script, document type, scan quality, and pipeline stage.
- Make every experiment reproducible from a saved configuration and dataset version.
System design
System modules
Corpus manager
Imports permitted documents, records language and layout metadata, and maintains fixed training, validation, and test splits.
Annotation workspace
Supports page transcription, field marking, relevant-passage selection, answer entry, and reviewer agreement checks.
OCR runner
Runs selected OCR engines through a common adapter and stores text, confidence, layout, latency, and failure information.
Retrieval runner
Indexes extracted text, executes prepared queries, and records ranked passages for lexical and semantic retrieval methods.
Question-answering runner
Produces answers from retrieved evidence and preserves the passages used for later scoring and review.
Evaluation dashboard
Calculates metrics, compares systems, filters errors, and exports tables and experiment records.
Methodology
System workflow
- 01Prepare corpus
Documents are checked for permitted use, classified, versioned, and assigned to fixed dataset splits.
- 02Create ground truth
Annotators transcribe selected pages and prepare fields, questions, answers, and relevant passages using written guidelines.
- 03Run pipelines
Selected OCR, retrieval, and question-answering configurations process the same frozen test set.
- 04Calculate metrics
The system scores each stage and groups results by language, document type, scan quality, and model.
- 05Analyse errors
Users inspect substitutions, missed fields, failed retrievals, unsupported answers, and pipeline-level error propagation.
Demonstration scenario
A prepared set of English, Hindi, and Kannada administrative documents is processed by two OCR configurations. The extracted text is indexed by two retrieval methods and used to answer a fixed question set. The dashboard compares stage-level metrics and traces one failed answer back to an OCR or retrieval error.
Engineering
Technical architecture
- Web application
- Nuxt and Vue for corpus review, annotation, experiment setup, result comparison, and error inspection.
- Evaluation API
- FastAPI for dataset access, model jobs, metric calculation, experiment state, and export endpoints.
- Data layer
- PostgreSQL for document metadata, annotations, dataset versions, experiment configurations, predictions, and scores.
- Model adapters
- Python interfaces that allow OCR, retrieval, and question-answering components to be replaced without changing the benchmark.
- Evaluation engine
- Stage-specific metrics with language-level grouping, confidence intervals, saved seeds, and fixed test splits.
- Reproducibility
- Docker-based setup with dataset manifests, configuration files, model version records, and exportable experiment results.
Testing
Evaluation
Evaluation measures
- Character and word error rates for OCR output
- Field extraction precision, recall, and F1 score
- Recall at k and mean reciprocal rank for retrieval
- Exact match, token F1, and evidence correctness for answers
- Performance differences across selected languages and document conditions
- Runtime, failure rate, and reproducibility across repeated experiments
System boundaries
- The corpus uses public, licensed, synthetic, or explicitly permitted documents only.
- The benchmark does not contain private student, citizen, or employee records.
- Results apply only to the documented dataset, languages, and model versions.
- The system evaluates document intelligence methods and does not make administrative decisions.
Included
- 01Document corpus and ground-truth management interface
- 02Replaceable OCR, retrieval, and question-answering pipelines
- 03Metric calculation and model-comparison dashboard
- 04Prepared benchmark dataset, experiments, and error analysis
- 05Complete source code in a private GitHub repository
- 06Complete project documentation with synopsis, abstract, methodology, processing diagrams, experiment results, screenshots, and conclusion
- 07Setup and usage guide
Project record
No buyer information is collected on this page.
- Permanent project ID
- GP-DA-1GYWHUI
- Catalogued
- 21 Aug 2026
- Completed
- Pending
- Verified
- Pending
- Demonstration
- Added when ready
Handover
After purchase
- 01Payment is confirmed
The project is marked unavailable and cannot be purchased again.
- 02Repository access is granted
The buyer's submitted GitHub account receives access to the private repository.
- 03The purchase record is delivered
The certification sheet is prepared from the reviewed buyer details and sent privately by email.