Building a Chatbot with LLM and Intent Recognition

Intent recognition once defined chatbot engineering, but the rise of large language models led many teams to abandon explicit classifiers for end-to-end prompting. That approach works for prototypes, yet production systems still benefit from a hybrid architecture: a fast intent router that guards a capable generation model. With dozens of inference providers now offering LLM access, the harder problem is not finding an API endpoint, but controlling cost and latency as conversational complexity grows. Oxlo.ai addresses this through request-based pricing and a broad model catalog that supports both classification and generation stages without cold starts.

Why Intent Recognition Persists in LLM Chatbots

Pure LLM chatbots treat every user message as an open-ended generation task. This creates three production burdens. First, latency increases because every response is generated from scratch rather than retrieved or routed. Second, cost scales with the full prompt length, including long system instructions and conversation history. Third, determinism suffers: the same user question can trigger different internal tool calls or safety policies on each invocation.

Explicit intent recognition solves these problems by collapsing the infinite input space into a finite set of known categories. Once the intent is identified, you can apply deterministic guardrails, query structured APIs, or select a specialized system prompt before invoking the large model. The result is lower latency, predictable behavior, and simpler audit trails.

Hybrid Architecture Overview

A production chatbot pipeline typically runs in two stages. The first stage classifies the user message into an intent. The second stage generates a contextual response, often using the intent to select a system prompt, a retrieval corpus, or a set of allowed tools. This separation of concerns lets you optimize each stage independently.

Oxlo.ai supports both stages from a single API key and OpenAI-compatible endpoint. For classification, you can use embedding models such as BGE-Large or E5-Large, or lightweight chat models such as DeepSeek V3.2. For generation, you can route to Llama 3.3 70B, Qwen 3 32B, Kimi K2.6, or DeepSeek R1 671B MoE depending on the required reasoning depth. Because Oxlo.ai offers 45-plus models across seven categories, you are not forced to over-provision a single large model for every subtask.

Implementing Intent Classification

Embedding-based classification is the fastest approach. You embed the user query and compare it against precomputed embeddings of intent descriptions using cosine similarity. The Oxlo.ai embeddings endpoint works as a drop-in replacement in the OpenAI SDK.

import openai

import numpy as np

client = openai.OpenAI(

    base_url="https://api.oxlo.ai/v1",

    api_key="YOUR_OXLO_API_KEY"

)

INTENTS = {

    "billing": "Questions about invoices, payments, or subscription changes",

    "technical": "Bug reports, integration errors, or API troubleshooting",

    "sales": "Pricing inquiries, feature comparisons, or