Dataset Preparation and Representation Scripts

This folder contains the official, verified dataset preparation, validation, and representation-building scripts for:

"Beyond Chunking: Node Integrity and Field Structure in Bengali Retrieval-Augmented Agent Memory"

1. Directory Structure

reproducibility/scripts/
├── README.md                           # This guide
├── data_ingestion/
│   ├── 00_fetch_external_data.py       # Validates external provenance and local file presence
│   ├── 01_prepare_corpora.py           # Standardizes knowledge corpora across Agri, Legal, and Edu
│   └── 02_prepare_queries.py           # Standardizes benchmark queries and evaluation subsets
├── validation/
│   └── 03_validate_data.py             # Independent reproducibility audit gate
├── representations/
│   ├── 04_build_l1_chunks.py           # L1 Chunked Unstructured builder (500-char window)
│   ├── 05_build_l2_full_nodes.py       # L2 Full-Node Unstructured builder (intact concatenated text)
│   └── 06_build_l3_fields.py           # L3 Structured BM25F schema builder (dedicated search slots)
├── 00_fetch_external_data.py           # Top-level entry point runner
├── 01_prepare_corpora.py               # Top-level entry point runner
├── 02_prepare_queries.py               # Top-level entry point runner
├── 03_validate_data.py                 # Top-level entry point runner
├── 04_build_l1_chunks.py               # Top-level entry point runner
├── 05_build_l2_full_nodes.py           # Top-level entry point runner
└── 06_build_l3_fields.py               # Top-level entry point runner

2. Script Execution Guide

Script 00: Check External Data Sources and Provenance

Validates external data repositories (HuggingFace datasets, NCTB curriculum textbooks, AIS/DAE handbooks):

python reproducibility/scripts/00_fetch_external_data.py

Script 01: Ingest and Standardize Corpora

Standardizes raw nodes into canonical JSON schema (structured_nodes.json) and exports selected_act_ids.json (735 legal acts) and nctb_provenance.json (190 secondary experiments):

python reproducibility/scripts/01_prepare_corpora.py

Script 02: Prepare Benchmark Queries and Evaluation Subsets

Standardizes 3,670 queries (queries.jsonl) and generates the 900-query agent sample, the 872-query common evaluated subset, the 28 timeout exclusions, and the 300-query perturbation sample:

python reproducibility/scripts/02_prepare_queries.py

Script 03: Execute Independent Data Validation Audit Gate

Executes the cryptographic and census audit verifying 100% hash integrity against manifest.json, node counts, query counts, chunk counts, and referential integrity:

python reproducibility/scripts/03_validate_data.py

Expected Output: DATA VALIDATION AUDIT: PASS

3. Representation Builders ($L_1$, $L_2$, $L_3$)

Script 04: Build / Verify $L_1$ (Chunked Unstructured)

Splits knowledge records into 500-character recursive windows with overlap (4,526 agri, 4,202 legal, 892 edu chunks):

python reproducibility/scripts/04_build_l1_chunks.py

Script 05: Build / Verify $L_2$ (Full-Node Unstructured)

Concatenates all active schema fields with newline delimiters into an intact document per record (1,028 agri, 735 legal, 190 edu intact documents):

python reproducibility/scripts/05_build_l2_full_nodes.py

Script 06: Build / Verify $L_3$ (Structured BM25F)

Extracts records into dedicated indexed field slots according to the official schema (Appendix Table 7), tagging designated query-aligned anchor fields:

python reproducibility/scripts/06_build_l3_fields.py

4. Design Principles

Standard Library Only: All scripts run in standard Python 3.8+ without external dependencies.
Console Encoding Safe: Automatic UTF-8 console reconfiguring prevents Windows cp1252 character errors.
Relocatable Execution: Scripts automatically detect project root from __file__ or accept explicit --data_dir and --output_dir arguments.