Documents
TXT, Markdown, PDF, DOCX and XLSX files are loaded from a local corpus and split into overlapping chunks.
PROTOTYPE · RAG · ISLAMIC FINANCE
Answer built on
found sources.
An educational AI agent takes a question on Islamic finance, finds relevant fragments in a local document corpus and returns an answer with source references, scores and chunk identifiers.
The task
Connect domain expertise with verifiable search
Questions about murabaha, riba or the functions of a Sharia committee require grounding in documents, not just a language model's memory. The prototype separates fact retrieval from answer generation and shows which fragments were used.
Pipeline
From document to answer with sources
FAISS + BM25
Semantic search via embeddings is combined with lexical search on exact terms.
Context selection
Optional CrossEncoder refines the candidate order before passing to the model.
GigaChat + sources
The model receives the selected fragments; the Gradio interface shows the answer, sources, score and chunk ID.
Implemented system
Not a notebook demo, a reproducible structure
Five formats
A unified ingestion layer for text, documents, tables and PDFs.
Persisted FAISS
Index and chunks are built by a separate ingest script and loaded by the application.
Gradio dialogue
Message history, example questions and a dedicated sources block.
Local .env
Keys are not committed and are excluded from Git; only a configuration template is public.
Modular package
Ingestion, embeddings, storage, retrieval and agent are split into components.
Basic checks
The project includes tests and reproducible run instructions.
Status and boundaries
Educational prototype, not a financial opinion
The system demonstrates RAG architecture and information retrieval. Answers require specialist verification, are not investment advice or a fatwa, and do not replace a Sharia board decision or legal expertise.