Case Study: Building the EnergiQ AI EV Assistant

In the rapidly evolving electric vehicle (EV) sector, user experience and data accessibility are critical. Drivers need quick answers about their charging status, transaction histories, and energy consumption, while also seeking instant support for technical charging issues.

The EnergiQ AI EV Assistant was built to solve these needs by delivering a voice-enabled, conversational agent integrated directly into the EnergiQ platform.

---

System Architecture & Workflow

The assistant functions as a highly integrated service within the FastAPI backend of the EnergiQ platform. It orchestrates user input processing, database interaction, document retrieval, and speech synthesis:

Request & Processing Flow

1. User Input: The user interacts with the system via voice (audio payload) or text.
2. Audio Transcription: If audio is received, the FastAPI backend routes it to OpenAI's Whisper-1 model for speech-to-text transcription.
3. Intent Routing: The LLM analyzes the query to determine the intent and required context.
4. Data Retrieval:
   • Database Queries: If requesting metrics, the agent executes Python tools via SQL function-calling against the PostgreSQL database.
   • Knowledge Queries: If asking general EV questions, the agent queries the vector database for document chunks using the RAG pipeline.
5. Response Generation: The agent constructs a final grounded response.
6. Voice Synthesis: If voice mode is active, the text is synthesized into audio using OpenAI's TTS-1 and streamed back to the user.

---

Core Engineering Implementations

1. Intent Routing & Function Calling Framework

To securely interface the agent with the user database, we established a robust function-calling mechanism using OpenAI’s tool definitions.

• Database Integration: The agent is equipped with specific tools such as:
  • getMonthlySpending(userId, year, month): Calculates the monthly charging costs.
  • getAvgSessionStats(userId): Returns average duration and energy charged.
  • getNearestStations(latitude, longitude, radius): Finds charging stations dynamically.
• Security & Constraints: To prevent unauthorized access, user context (e.g., userId) is injected on the backend server side and is not exposed to client-side input parameters.
• Performance: Evaluation on sample queries yielded an Intent Accuracy F1-score of 0.94, ensuring users get correct information without wrong database queries.

2. High-Recall RAG (Retrieval-Augmented Generation) Pipeline

For general knowledge inquiries (e.g., "What is the difference between a CCS and a CHAdeMO connector?"), we implemented a LangChain-powered RAG pipeline:

• Document Processing: Internal documentation, platform FAQs, and EV specs were chunked and embedded using OpenAI's text embeddings.
• Contextual Retrieval: During runtime, a custom retriever fetches relevant document chunks. The prompt is dynamically populated with these chunks to ground the LLM's answers.
• Metrics: The retrieval configuration was fine-tuned to achieve a Recall@3 of 0.95, virtually eliminating hallucinations for platform-related FAQs.

3. Latency-Optimized Voice Chat Endpoint

A core feature of the assistant is the low-latency /voice endpoint, allowing full speech-to-speech interaction.

• Speech-to-Text: Incoming WAV/MP3 files are routed to OpenAI's whisper-1 engine.
• Text-to-Speech: The generated textual response is synthesized into natural-sounding speech using OpenAI's tts-1 model.
• Inference Optimization: By optimizing payload sizes, connection reuse, and response stream formatting, the end-to-end token generation latency was minimized to 78 ms/token, enabling an interactive voice conversational flow.

---

Deployment & Scale

The backend is written in FastAPI, using SQLAlchemy ORM for PostgreSQL queries, and deployed on Railway. The service is configured to handle scaling loads targeting 7,000 active users across the EnergiQ EV management platform.