Management Document and Stock Data of Nine Indonesian Companies for AI Research

GENERAL OVERVIEW This dataset contains 27 official reports from 9 companies listed on the Indonesia Stock Exchange (IDX) for the 2024 fiscal year, along with their daily stock price data. These reports were processed into plain text and search indices, which are used by a Retrieval-Augmented Generation (RAG) system to answer queries regarding the companies. Ticker Company Name Sector BBCA Bank Central Asia Banking BMRI Bank Mandiri Banking BBRI Bank Rakyat Indonesia Banking BBNI Bank Negara Indonesia Banking TLKM Telkom Indonesia Telecommunications ASII Astra International Automotive JSMR Jasa Marga Toll Roads UNVR Unilever Indonesia Consumer Goods ANTM Aneka Tambang Mining 1. RAW DATA (Unprocessed Source Data) dokumen_pdf/ Contents: 27 PDF files grouped by company. Each company includes three documents: Annual Report Sustainability Report (ESG) Financial Report Source: Downloaded on March 17, 2026, from the official Investor Relations webpage of each company. Three ANTM documents were retrieved from the IDX website (idx.co.id), Antam's official CDN, and IndoPremier. Full URLs for each file are listed in metadata_dokumen.csv. Purpose: Primary knowledge base for the RAG system. metadata_dokumen.csv Contents: Catalog of all 27 documents containing company name, document type, publication year, language, page count, source URL, download timestamp, and the SHA-256 checksum for each file. Source: Compiled manually by the researcher during document collection and inspection. Purpose: Corpus registry and data provenance tracking. SHA-256 hashes ensure file integrity. harga_saham_mentah.csv Contents: Daily historical stock prices for the 9 companies (Open, High, Low, Close, Adjusted Close, Volume) covering January 2, 2020, to March 16, 2026. Source: Yahoo Finance (tickers suffixed with .JK, e.g., BBCA.JK). Purpose: Supplementary dataset for stock price analysis features. 2. PREPROCESSED DATA (Sequential Pipeline: Steps 1 to 5) 1_teks_hasil_ekstraksi/ Contents: Raw extracted text from each PDF (27 .txt files), retaining raw formatting, page breaks, headers, and footers. Source: Text extraction stage from dokumen_pdf/. Purpose: Intermediate audit trail. This is the only stage that preserves original page numbers. 2_teks_bersih/ Contents: Cleaned text where recurring headers and footers have been stripped, hyphenated line breaks reconnected, and character encodings normalized. Source: Text cleaning pipeline applied to 1_teks_hasil_ekstraksi/. Purpose: Normalized input corpus for subsequent text chunking. 3_chunks/all_chunks.jsonl Contents: 5,753 text chunks (1,000 words per chunk with a 150-word sliding overlap). Each chunk includes metadata attributes: company ticker, document type, and fiscal year. Source: Document chunking applied to 2_teks_bersih/. Purpose: Retrieval unit for semantic search and grounded citation generation in RAG responses. 4_index_faiss/ Contents: Semantic search index files: faiss.index: Dense vector search index. embeddings.npy: Dense vector embeddings for each chunk. id_map.json: Mapping layer connecting vector IDs to source chunks. embedding_model_name.txt: Embedding model identifier (all-MiniLM-L6-v2). build_meta.json: Build timestamps and index hyperparameters. Source: Generated from 3_chunks/ using the all-MiniLM-L6-v2 embedding model. Purpose: Low-latency vector retrieval (sub-0.1s response time) for user queries. 5_data_terstruktur/ metadata_dokumen_bersih.parquet Contents: Parquet version of metadata_dokumen.csv with optimized schema typing. Source: Direct schema cast and conversion from metadata_dokumen.csv. Purpose: High-performance, production-ready metadata access. harga_saham_bersih.parquet Contents: Price series enriched with engineered features: daily_return: Percentage daily return. return_outlier: Flag for extreme price movement events. volume_outlier: Flag for anomalous trading volume spikes. Source: Feature engineering applied to harga_saham_mentah.csv. Purpose: Analytics-ready financial dataset. 3. PROCESSING PIPELINE PDF Documents (dokumen_pdf/) │ ▼ [Text Extraction] Raw Extracted Text (1_teks_hasil_ekstraksi/) │ ▼ [Cleaning & Normalization] Cleaned Text (2_teks_bersih/) │ ▼ [1,000-Word Chunking, 150-Word Overlap] Text Chunks (3_chunks/all_chunks.jsonl) │ ▼ [Vector Embeddings (all-MiniLM-L6-v2)] FAISS Vector Index (4_index_faiss/) │ ▼ RAG System Pipeline: User Query ──▶ Vector Search ──▶ Contextual Answer + Citations Raw Stock Data (harga_saham_mentah.csv) │ ▼ [Feature Engineering] Cleaned Stock Data (5_data_terstruktur/harga_saham_bersih.parquet) 4. NOTES Reporting Year Discrepancy: The BCA Sustainability Report is dated 2025, whereas all other documents correspond to fiscal year 2024. Language Distribution: Three documents are written in Indonesian (ANTM Annual Report, ANTM Sustainability Report, and UNVR Financial Report); the remaining 24 documents are in English. Structural Heterogeneity: Financial reports vary substantially in format, ranging from multi-hundred-page complete audited statements to concise executive summaries spanning fewer than 20 pages. Chunking Strategy: The chunks in this directory use a fixed 1,000-word window (the baseline setup from the paper). The optimized configuration (254 word-pieces) is not included in this directory.