Gold Price Conversational AI Assistant

01. The Real Problem

LLMs excel at linguistic synthesis, but when queried about financial commodities like gold, baseline generative models frequently hallucinate rates, extrapolate speculative trends, and conflate historical spot prices with live market data. Furthermore, coupling speculative conversational text generation with financial checkout workflows in a single endpoint creates catastrophic failure modes in state handling and security.

The goal of this project was to construct a disciplined, self-contained Python backend that:

• Grounds conversational answers in historical tabular price records through vector similarity search rather than LLM parametric memory.
• Categorizes incoming intent (e.g., date-specific queries, 30-day rolling trend evaluations, or lowest price months) using deterministic logic before invoking the LLM.
• Strictly isolates conversational exploration (API 1) from transaction recording and settlement (API 2).

02. Verified System Architecture

Inspection of the repository's source code (specifically app/main.py, app/model.py, and app/purchase.py) confirms a dual-service architecture designed around separation of concerns:

03. RAG & Tabular Retrieval Pipeline

In app/model.py and app/utils.py, historical market data from Gold Price.csv is cleaned and transformed into semantic text summaries to make tabular numerical records understandable to dense vector embeddings:

def default_sentence(row):
    open_ = int(row['open']) if 'open' in row and pd.notna(row['open']) else None
    high  = int(row['high']) if 'high' in row and pd.notna(row['high']) else None
    low   = int(row['low'])  if 'low'  in row and pd.notna(row['low'])  else None
    close = int(row['price']) if pd.notna(row['price']) else None

    bits = [f"On {row['date'].strftime('%Y-%m-%d')}"]
    if open_ is not None: bits.append(f"opened at ₹{open_}")
    if high  is not None: bits.append(f"peaked at ₹{high}")
    if low   is not None: bits.append(f"dropped to ₹{low}")
    if close is not None: bits.append(f"closed at ₹{close}")
    return ", ".join(bits) + "."

Sentences are embedded into ChromaDB with all-MiniLM-L6-v2. When querying, the retrieval engine queries the nearest 5 neighbors and enforces a confidence distance threshold of distances[0] < 0.8 before feeding context to DeepSeek R1.

04. Intent Classification & Routing

Instead of passing all queries unconditionally to the generative model, classify_query() parses the prompt into distinct routing categories:

Furthermore, is_purchase_intent() scans for conversion intent (e.g. buy, purchase, invest, ₹, grams), prompting the assistant to deliver structured next_action payload directing the user to the transaction interface.

05. Decoupled Purchase Engine (API 2)

In app/purchase.py, a completely separate FastAPI application provides a dedicated financial processing endpoint:

• Fixed-Point Decimal Arithmetic: Uses Python's standard decimal.Decimal and ROUND_DOWN to eliminate floating-point rounding errors common in currency conversions.
• Thread Safety: Uses threading.Lock() when appending to the local transaction CSV storage to prevent race conditions during concurrent write operations.
• UUID & Idempotency: Generates unique transaction IDs (uuid4) and ISO 8601 UTC timestamps for every completed order.

Authentic Interface Artifacts

06. Engineering Decisions & Honest Constraints