Remove nested folder; keep Space files at repo root

Capstone-Project-i/README.md

@@ -1,43 +0,0 @@
----
-title: Regulatory Loan Analyst (FDIC 3.2)
-emoji: 🏦
-colorFrom: blue
-colorTo: green
-sdk: gradio
-sdk_version: "5.12.0"
-app_file: app.py
-pinned: false
----
-
-
-
-## 📄 Regulatory Loan Evaluation Assistant
-
-This application is a **prompt-engineered regulatory reasoning system** designed for
-loan evaluation in accordance with the **FDIC RMS Manual of Examination Policies – Section 3.2 (Loans)**.
-
-### 🔍 What this system does
-- Extracts **structured loan facts** from uploaded loan documents using OCR
-- Answers user questions using **only**:
-  - Extracted loan facts, and
-  - FDIC Section 3.2 regulatory guidance
-- Refuses non-loan or out-of-scope questions
-- Avoids approvals, rejections, or predictions
-
-### 🧠 Key Design Principles
-- **Prompt engineering only** (no model training or fine-tuning)
-- **Single source of truth** for regulatory reasoning
-- **Audit-ready**, document-grounded responses
-- **Regulatory tone** aligned with examiner expectations
-
-### 📌 Inputs
-- Optional loan documents (PDF / image)
-- User regulatory or loan-related questions
-
-### 🚫 Explicitly excluded
-- Credit scoring
-- Automated decisions
-- OCR beyond basic text extraction
-- External data sources
-
-This project is intended for **educational and regulatory analysis purposes only**.

Capstone-Project-i/app.py

@@ -1,334 +0,0 @@
-#!/usr/bin/env python
-# coding: utf-8
-
-# =========================================================
-# 1. IMPORTS & ENV
-# =========================================================
-import os
-import json
-import re
-import hashlib
-from dotenv import load_dotenv
-from PIL import Image
-
-import gradio as gr
-import pytesseract
-from pdf2image import convert_from_path
-from groq import Groq
-
-load_dotenv()
-
-# =========================================================
-# 2. LOAD FDIC SECTION 3.2 ONCE (GLOBAL)
-# =========================================================
-with open("data/fdic_section_3_2_chunks_refined.json") as f:
-    FDIC_CHUNKS = json.load(f)
-
-# =========================================================
-# 3. GROQ CLIENT & MODELS
-# =========================================================
-client = Groq(api_key=os.environ.get("GROQ_API_KEY"))
-
-MODEL_LLM1 = "llama-3.1-8b-instant"   # OCR → loan summary
-MODEL_LLM2 = "llama-3.1-8b-instant"   # topic indexing
-MODEL_LLM4 = "meta-llama/llama-4-scout-17b-16e-instruct"  # reasoning
-
-# =========================================================
-# 4. SESSION STATE
-# =========================================================
-SESSION_STATE = {
-    "ocr_text": "",
-    "loan_summary": None
-}
-
-OCR_CACHE = {}
-
-# =========================================================
-# 5. GUARDRAILS
-# =========================================================
-NON_LOAN_KEYWORDS = [
-    "movie", "music", "sports", "weather", "joke", "recipe",
-    "health", "cold", "fever", "doctor", "medicine",
-    "politics", "election"
-]
-
-def sanitize_user_input(text):
-    return text.strip()[:5000] if text else ""
-
-def is_non_loan_question(text):
-    return any(k in text.lower() for k in NON_LOAN_KEYWORDS)
-
-# =========================================================
-# 6. SAFE JSON PARSER
-# =========================================================
-def safe_json_loads(text, stage):
-    if not text:
-        raise ValueError(f"{stage} returned empty response")
-
-    text = re.sub(r"```json|```", "", text).strip()
-    match = re.search(r"\{.*\}", text, re.DOTALL)
-
-    if not match:
-        raise ValueError(f"{stage} returned no JSON:\n{text}")
-
-    return json.loads(match.group())
-
-# =========================================================
-# 7. OCR HELPERS
-# =========================================================
-MAX_PAGES = 5
-
-def file_hash(path, max_bytes=1024 * 1024):
-    h = hashlib.md5()
-    with open(path, "rb") as f:
-        h.update(f.read(max_bytes))
-    return h.hexdigest()
-
-def ocr_file(path):
-    if path.lower().endswith(".pdf"):
-        text = ""
-        pages = convert_from_path(path, dpi=200)[:MAX_PAGES]
-        for p in pages:
-            text += pytesseract.image_to_string(p.convert("L")) + "\n"
-        return text.strip()
-    else:
-        img = Image.open(path).convert("L")
-        return pytesseract.image_to_string(img).strip()
-
-def run_ocr_pipeline(uploaded_files):
-    texts = []
-    for f in uploaded_files:
-        path = str(f)
-        key = file_hash(path)
-        if key not in OCR_CACHE:
-            OCR_CACHE[key] = ocr_file(path)
-        texts.append(OCR_CACHE[key])
-    return "\n".join(texts)
-
-# =========================================================
-# 8. LOAN SCHEMA
-# =========================================================
-LOAN_SCHEMA = """<same as your original schema>"""
-
-# =========================================================
-# 9. SYSTEM PROMPTS
-# =========================================================
-LLM1_SYSTEM_PROMPT = f"""
-You are an information extraction engine for bank loan documents.
-
-Task:
-- Extract ONLY facts that are explicitly stated in the text.
-- Do NOT infer, assume, normalize, or calculate anything.
-- If a value is missing or unclear, use null or "unknown".
-
-Rules:
-- Use ONLY the provided OCR text.
-- Do NOT add explanations.
-- Do NOT reference regulations.
-- Output MUST strictly match the schema below.
-- Return ONLY valid JSON.
-
-Schema:
-{LOAN_SCHEMA}
-"""
-
-LLM2_SYSTEM_PROMPT = """
-You are a regulatory topic indexing assistant.
-
-Inputs:
-- A user question
-- A list of FDIC RMS Manual Section 3.2 headings with chunk_ids
-
-Task:
-- Select ONLY the chunk_ids whose headings are directly relevant
-  to answering the user question.
-- Base your decision ONLY on the heading titles.
-- Do NOT interpret or summarize policy text.
-
-Rules:
-- Select between 1 and 6 chunk_ids.
-- If no headings are relevant, return an empty list.
-- Do NOT explain your reasoning.
-- Return ONLY valid JSON.
-
-Output format:
-{
-  "selected_chunk_ids": ["string"]
-}
-"""
-
-LLM4_SYSTEM_PROMPT = """
-You are a regulatory-aligned loan evaluation assistant.
-
-You are given TWO authoritative sources:
-
-SOURCE A — Loan Summary
-• Structured facts extracted from uploaded loan documents
-• This is the ONLY source for borrower name, loan type, interest rate,
-  amounts, collateral, and other loan-specific details
-
-SOURCE B — FDIC RMS Manual Section 3.2 (Loans)
-• This is the ONLY source for regulatory objectives, examiner expectations,
-  loan review systems, risk management, and policy intent
-
-RULES (STRICT):
-1. If the user asks for loan details → answer ONLY from SOURCE A
-2. If the user asks regulatory or examiner questions → answer ONLY from SOURCE B
-3. If the user asks a mixed question → clearly separate:
-   • factual loan details (SOURCE A)
-   • regulatory interpretation (SOURCE B)
-4. Do NOT infer or assume missing facts
-5. Do NOT use general banking knowledge
-6. Do NOT approve, reject, or predict loan outcomes
-7. If required information is missing, explicitly state that it is not available
-
-Tone:
-Professional, neutral, examiner-style.
-No markdown. No speculation.
-
-"""
-
-NO_DOC_PROMPT = f"""
-You are creating a placeholder loan summary.
-
-Rules:
-- Use ONLY the schema provided.
-- Do NOT infer or fabricate details.
-- Populate fields only if explicitly stated in the user input.
-- Otherwise, use null or "unknown".
-- Return ONLY valid JSON.
-
-Schema:
-{LOAN_SCHEMA}
-"""
-
-# =========================================================
-# 10. LLM CALL
-# =========================================================
-def call_llm(system_prompt, user_prompt, model, temperature=0):
-    r = client.chat.completions.create(
-        model=model,
-        temperature=temperature,
-        messages=[
-            {"role": "system", "content": system_prompt},
-            {"role": "user", "content": user_prompt}
-        ]
-    )
-    return r.choices[0].message.content.strip()
-
-# =========================================================
-# 11. MAIN LOGIC (FINAL)
-# =========================================================
-def process_request(user_text, uploaded_files):
-    user_text = sanitize_user_input(user_text)
-
-    # 🚫 NON-LOAN GUARDRAIL
-    if is_non_loan_question(user_text):
-        return "", "⚠️ Only FDIC Section 3.2 loan and regulatory questions are supported."
-
-    # ======================================================
-    # LLM-1: OCR → Loan Summary (ONLY if files exist)
-    # ======================================================
-    if uploaded_files:
-        ocr_text = run_ocr_pipeline(uploaded_files)
-
-        loan_summary = safe_json_loads(
-            call_llm(
-                LLM1_SYSTEM_PROMPT,
-                ocr_text,
-                MODEL_LLM1
-            ),
-            "LLM-1"
-        )
-
-        SESSION_STATE["ocr_text"] = ocr_text
-        SESSION_STATE["loan_summary"] = loan_summary
-
-    else:
-        # Follow-up or regulatory-only question
-        ocr_text = SESSION_STATE.get("ocr_text", "")
-        loan_summary = SESSION_STATE.get("loan_summary")
-
-        # ❗ Do NOT force NO-DOC extraction for regulatory questions
-        if loan_summary is None:
-            loan_summary = {
-                "note": "No loan documents uploaded. Loan-specific facts unavailable."
-            }
-
-    # ======================================================
-    # LLM-2: FDIC Section 3.2 Topic Indexing (HEADINGS ONLY)
-    # ======================================================
-    headings_payload = {
-        "user_question": user_text,
-        "fdic_headings": [
-            {
-                "chunk_id": c["chunk_id"],
-                "heading": c.get("subtopic") or c.get("title")
-            }
-            for c in FDIC_CHUNKS
-        ]
-    }
-
-    selected_ids = safe_json_loads(
-        call_llm(
-            LLM2_SYSTEM_PROMPT,
-            json.dumps(headings_payload),
-            MODEL_LLM2
-        ),
-        "LLM-2"
-    ).get("selected_chunk_ids", [])
-
-    selected_chunks = [
-    {
-        "chunk_id": c["chunk_id"],
-        "heading": c.get("subtopic") or c.get("title")
-    }
-    for c in FDIC_CHUNKS
-    if c["chunk_id"] in set(selected_ids)
-][:6]   # 🔒 HARD CAP (very important)
-
-
-    # ======================================================
-    # LLM-4: FINAL REGULATORY + FACTUAL ANSWER
-    # ======================================================
-    llm4_payload = {
-        "loan_summary": loan_summary,
-        "fdic_section_3_2": selected_chunks,
-        "user_question": user_text
-    }
-
-    answer = call_llm(
-        LLM4_SYSTEM_PROMPT,
-        json.dumps(llm4_payload),
-        MODEL_LLM4,
-        temperature=0.2
-    )
-
-    return ocr_text, answer
-
-# =========================================================
-# 12. GRADIO UI
-# =========================================================
-def chat_handler(user_text, uploaded_files, chat_history):
-    chat_history = chat_history or []
-    _, answer = process_request(user_text, uploaded_files)
-    chat_history.append({"role": "user", "content": user_text})
-    chat_history.append({"role": "assistant", "content": answer})
-    return chat_history
-
-with gr.Blocks(title="Regulatory Loan Evaluation Assistant") as demo:
-    gr.Markdown("## 📄 Regulatory Loan Evaluation Assistant")
-    chat = gr.Chatbot(height=450)
-    files = gr.File(
-        label="Upload Loan Documents (Optional)",
-        file_types=[".pdf", ".png", ".jpg", ".jpeg"],
-        file_count="multiple"
-    )
-    user_input = gr.Textbox(placeholder="Ask a regulatory or loan question")
-    gr.Button("Send").click(
-        fn=chat_handler,
-        inputs=[user_input, files, chat],
-        outputs=[chat]
-    )
-
-demo.launch()

Capstone-Project-i/packages.txt

@@ -1,2 +0,0 @@
-tesseract-ocr
-poppler-utils

Capstone-Project-i/requirements.txt

@@ -1,7 +0,0 @@
-gradio>=4.0
-pytesseract
-pdf2image
-pillow
-python-dotenv
-groq
-opencv-python-headless