RaiyanKhaan/Beyond-Chunking

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

Official anonymized reproduction suite for the paper:

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

---

Overview

This repository provides complete, self-contained artifacts, datasets, experimental code, and evaluation oracles to independently verify all empirical claims, tables, and mechanistic analyses presented in the paper.

The suite evaluates a three-level representation ladder across three Bengali domains:

1. $L_1$ (Chunked Unstructured): Source documents split into fixed 500-character windows.
2. $L_2$ (Full-Node Unstructured): Coherent records kept intact as single documents without field demarcation.
3. $L_3$ (Structured BM25F): Intact records indexed into distinct functional field slots (uniform unit weighting $M_3^u$ and benchmark-fitted ceiling $M_3^t$).

---

Repository Structure

reproducibility/
├── README.md                      # This master reproduction guide
├── requirements.txt               # Pinned Python package dependencies
├── LICENSE-CODE                   # MIT License for source code
├── LICENSE-DATA.md                # CC-BY-4.0 licensing declaration for benchmark data
├── verify_paper_numbers.py        # Automated paper-number oracle audit gate (< 2s execution)
├── reproduce_all.py               # Master end-to-end headless CLI reproduction runner
│
├── dataset/                       # Benchmark data & cryptographic checksums
│   ├── manifest.json              # SHA-256 hashes and census for all 20 benchmark files
│   ├── DATA_SOURCES.md            # Detailed provenance, licenses, and collection protocols
│   ├── processed/                 # Fully processed benchmark partitions (Agri, Legal, Edu)
│   │   ├── agriculture/           # 1,028 nodes, 4,526 chunks, 2,000 queries, gold mappings
│   │   ├── legal/                 # 735 acts, 4,202 chunks, 1,100 queries, gold mappings
│   │   └── education/             # 190 experiments, 892 chunks, 570 queries, gold mappings
│   └── derived/                   # Standardized evaluation subsets
│       ├── agent_subsets/         # 872 common subset + 28 documented network timeouts
│       └── perturbation_subsets/  # 300-query counterfactual perturbation sample
│
├── notebooks/                     # Interactive Jupyter Notebooks with cached cell outputs
│   ├── 01_agriculture_experiments.ipynb  # Full agricultural advisory pipeline
│   ├── 02_legal_experiments.ipynb        # Bangladesh statutory law experiments
│   ├── 03_education_experiments.ipynb    # Secondary science procedural experiments
│   └── 04_agent_memory_experiments.ipynb # ReAct agent memory & perturbation evaluations
│
├── scripts/                       # Modular headless CLI execution pipeline
│   ├── 03_validate_data.py        # Cryptographic hash and referential integrity audit gate
│   ├── 07_run_core_bm25.py        # Reproduces Table 2 (L1 vs L2 vs L3 flat)
│   ├── 08_run_uniform_bm25f.py    # Reproduces Table 2 (L3 uniform BM25F)
│   ├── 09_tune_bm25f_ceiling.py   # Reproduces Table 10 & Table 11 (Tuned ceiling)
│   ├── 10_diagnose_fragmentation.py # Reproduces Table 3 (Diagnostic failure attribution)
│   ├── 13_field_shuffle.py        # Reproduces Table 4 (Counterfactual field shuffles)
│   ├── 14_run_agent.py            # Reproduces Table 5 (Agent memory benchmark)
│   ├── 15_agent_anchor_shuffle.py # Reproduces Table 6 (Counterfactual agent grounding)
│   ├── 20_evaluate_all.py         # Executes all evaluation scripts sequentially
│   └── 23_verify_paper_numbers.py # Top-level CLI entrypoint for the verification oracle
│
├── src/                           # Modular core library
│   ├── memory_backends.py         # Unified memory backends (BM25Okapi, FullNode, BM25F)
│   └── agent_runner.py            # ReAct agent runtime with token accounting and logging
│
└── outputs/                       # Precompiled empirical results matching paper tables
    ├── retrieval/                 # Standalone sparse retrieval metrics (Tables 2, 8, 10)
    ├── failure_diagnosis/         # Diagnostic attribution partitions (Table 3)
    ├── counterfactuals/           # Field shuffle results (Table 4)
    └── agent_memory/              # Agent recall, citation grounding, and traces (Tables 5, 6)

---

Quickstart (Under 2 Minutes)

1. Environment Setup

# Clone the repository
git clone <anonymous-github-repo-url>
cd reproducibility

# Create and activate a virtual environment
python -m venv venv
source venv/bin/activate        # On Linux/macOS
# .\venv\Scripts\activate       # On Windows

# Install dependencies
pip install -r requirements.txt

2. Run the 10-Second Paper Verification Oracle

To verify that all empirical numbers reported in the paper's tables and prose match the experimental codebase exactly, run:

python verify_paper_numbers.py

This script asserts every metric in the paper against saved evaluation outputs, verifying:

• Table 1: Benchmark node, chunk, and query census (1,953 nodes, 9,620 chunks, 3,670 queries).
• Table 2: Representation ladder Recall@10 across all 3 domains ($L_1 \to L_2 \to L_3$).
• Table 3: Fragmentation failure recovery categorization ($n=731$ recovered queries).
• Table 4: Counterfactual perturbation controls (Unshuffled reference, Full, Content, Identity).
• Table 5: Agent memory performance on the 872-query common subset ($M_1, M_2, M_3^u, M_3^t$).
• Table 6: Counterfactual agent memory grounding ($N=300$).
• Tables 8, 10, 12: Standalone ablations, tuned BM25F ceilings, and dense model baselines.

3. Run the Cryptographic Data Integrity Gate

To independently audit the integrity of the corpus, referential query-to-node mappings, and SHA-256 hashes of all 20 registered files:

python scripts/03_validate_data.py

Expected output: ```text

DATA VALIDATION AUDIT: PASS All experimental datasets, hashes, subsets, and mappings are verified.

---

## Paper Table to Code Mapping

Every table in the paper corresponds directly to an automated CLI script, a precomputed JSON artifact, and an interactive notebook:

| Paper Table | Description | Headless CLI Script | Interactive Notebook | Saved Artifact |
|:---|:---|:---|:---|:---|
| **Table 1** | Corpus and Query Census | `scripts/03_validate_data.py` | `01_agriculture`, `02_legal`, `03_education` | `dataset/manifest.json` |
| **Table 2** | Representation Ladder (R@10) | `scripts/07_run_core_bm25.py`<br>`scripts/08_run_uniform_bm25f.py` | `01_agriculture`, `02_legal`, `03_education` | `outputs/retrieval/core_bm25_results.json`<br>`outputs/retrieval/uniform_bm25f_results.json` |
| **Table 3** | Fragmentation Failure Diagnosis | `scripts/10_diagnose_fragmentation.py` | `01_agriculture`, `02_legal`, `03_education` | `outputs/failure_diagnosis/l1_l2_diagnosis_summary.json` |
| **Table 4** | Counterfactual Field Shuffles | `scripts/13_field_shuffle.py` | `01_agriculture`, `02_legal`, `03_education` | `outputs/counterfactuals/table4_field_shuffle_results.json` |
| **Table 5** | Agent Memory Performance ($n=872$) | `scripts/14_run_agent.py` | `04_agent_memory_experiments.ipynb` | `outputs/agent_memory/table5_agent_results.json` |
| **Table 6** | Agent Memory Permutations ($N=300$) | `scripts/15_agent_anchor_shuffle.py` | `04_agent_memory_experiments.ipynb` | `outputs/agent_memory/table6_agent_perturbation_results.json` |
| **Table 8** | Standalone Field-Group Ablations | Notebook cell executions | `01_agriculture`, `02_legal`, `03_education` | `outputs/retrieval/ablation_group_metrics.json` |
| **Table 10** | Tuned BM25F Ceilings ($L_3^t$) | `scripts/09_tune_bm25f_ceiling.py` | `01_agriculture`, `02_legal`, `03_education` | `outputs/retrieval/tuned_ceiling_results.json` |
| **Table 12** | Dense & Hybrid Robustness Checks | Notebook cell executions | `01_agriculture`, `02_legal`, `03_education` | `outputs/retrieval/dense_hybrid_results.json` |

---

## Interactive Notebooks

All four notebooks in `notebooks/` are pre-executed with cell outputs retained, enabling reviewers to inspect results immediately without re-running long dense model or LLM inference loops:

* **`notebooks/01_agriculture_experiments.ipynb`:** Evaluates 1,028 agricultural nodes across Dhaka standard and 16 regional dialects; reproduces Tables 2, 3, 4, 8, 10, and 12.
* **`notebooks/02_legal_experiments.ipynb`:** Evaluates 735 statutory acts and 1,100 citizen legal inquiries; demonstrates the legal counterexample where uniform fielding trails intact full nodes ($77.82\%$ vs. $79.55\%$).
* **`notebooks/03_education_experiments.ipynb`:** Evaluates 190 secondary science lab experiments; isolates the $+52.46$pp anti-fragmentation recovery.
* **`notebooks/04_agent_memory_experiments.ipynb`:** Evaluates the ReAct agent (`gemini-2.5-flash-lite`, 1 tool, at most 2 calls) across memory conditions $M_1, M_2, M_3^u, M_3^t$ and counterfactual anchor shuffles.

Each notebook features automatic environment detection: works seamlessly locally, on Google Colab, or in Kaggle notebook environments.

---

## Data Provenance and Licensing

* **Source Materials:**
  * **Agriculture:** Derived from the Agricultural Information Service (AIS) and Directorate of Agricultural Extension (DAE), Ministry of Agriculture, Bangladesh.
  * **Legal:** Derived from official statutory enactments published under the Bangladesh Code, Legislative and Parliamentary Affairs Division, Ministry of Law.
  * **Education:** Derived from secondary school science laboratory curricula published by the National Curriculum and Textbook Board (NCTB).
* **Licenses:**
  * Source code is released under the **MIT License** ([LICENSE-CODE](LICENSE-CODE)).
  * Curated benchmark records, queries, and evaluation splits are licensed under **Creative Commons Attribution 4.0 International (CC-BY-4.0)** ([LICENSE-DATA.md](LICENSE-DATA.md)).
  * For full provenance and reconstruction instructions, see [`dataset/DATA_SOURCES.md`](dataset/DATA_SOURCES.md).

---

## Anonymity and Compliance

This repository is submitted in full compliance with the double-blind review policy:
* All author identities, personal emails, institutional affiliations, and local development paths have been scrubbed.
* All evaluation numbers in the paper correspond strictly to frozen benchmark configurations without post-hoc data alterations.