Design Guide

---

Xena SignalX Chat Based AI Service - Design

Overview

Xena is a chat-based query interface for platforms RM and Dart, supporting risk-based, view-based, and data-based queries using LLM-based reasoning, vector database search, and classification mechanisms.

• Risk-Based Query Answering: LLM-driven responses within due diligence, customizable across risk domains.
• View-Based Query Answering: Uses precomputed embeddings for structured responses.
• Data-Based Query Answering: Dynamic embeddings with vector search for RM & Dart platforms.

---

Core Functional Requirements

Platform	Query Type	Storage
RM	View-Based	Milvus Lite (Temporary)
RM	Risk-Based	LLM Only
Dart	View-Based	Milvus Lite (Temporary)
Dart	Data-Based	Milvus Standalone (Persistent)

---

Modules

1. Query Classification: Routes queries into view, data, or risk categories for both platforms.
2. LLM-Based Query Answering: Single risk-focused LLM, using domain-specific prompts.
3. View-Based Query Handling: Fetches from Milvus Lite, calls LLM for structured response.
4. Data-Based Query (Dart): Uses Milvus Standalone for persistent vector search accuracy.
5. Efficient Searching: Metadata filtering for fast, relevant vector retrieval (Dart data queries).

---

Architecture Overview

• Modular microservices, shared across RM & Dart.
• Dockerized and orchestrated via Kubernetes.
• Milvus Lite for temporary embeddings; Milvus Standalone for persistent embeddings.
• WebSockets supported for real-time query responses.

---

Processing Flow

1. User query enters system
2. Classification: Categorize as view, data, or risk-based
3. Processing:
   • View: Fetch from Milvus Lite → Call LLM → Return response
   • Data: Vector search in Milvus Standalone → Call LLM → Return response
   • Risk: Direct LLM call → Domain-specific prompt → Return response

---

Deployment Strategy

• Microservices: Each query type (classification, view, data, risk) independently deployable
• API Gateway: Routes via REST or WebSockets

---

Low-Level Design (LLD)

Modules and Functions

1. Query Classification

   • Keyword/LLM-based logic
   • Determines query type (view/data/risk)

2. View-Based Query Processing

   • Embedding fetches from Milvus Lite
   • LLM generates responses from context and embeddings

3. Data-Based Query Processing (Dart Only)

   • Vector search via Milvus Standalone
   • LLM response generated using results and query

4. Risk-Based Query Processing

   • Direct LLM response engineered for risk domain

5. Efficient Searching (Dart Data-Based)

   • Metadata filtering for fast, relevant vector retrieval

---

High-Level Component Overview (v0.1.0)

• FastAPI backend: Handles requests, endpoints
• Milvus vector database: Embedding storage/retrieval
• SentenceTransformer model: Generates embeddings
• Authentication: API security
• Services: Business logic implementation

---

Data Flow

1. Document ingestion
2. Embedding generation
3. Storage in Milvus
4. Query processed to embedding
5. Semantic search for similarity
6. Response returned via API

---

API Modules

• API Server (server.py): FastAPI endpoints/routes, middleware
• Embeddings (embeddings.py): MiniLM-L6-v2 for embeddings
• Milvus Client (milvusclient.py): Connects with DB, vector search
• Document Processing (loadview.py): Parses, stores docs
• Authentication (authutils.py): Middleware
• Database (dbservice.py): Data persistence
• Business Logic: Query routing, suggestion, response handling

---

Deployment

Prerequisites

# Start Milvus Database
docker-compose up -d

# Install dependencies
poetry install

# Run FastAPI Server
uvicorn app.serverapp --host 0.0.0.0 --port 8000