Blog post

Free

Building a mostly autonomous HR assistant with ChatGPT and LangChain agents

FreeFree tier

Inputs: textOutputs: text

Type

Open Source

Visit Website GitHub (40 stars)

About Blog post

This article from Towards AI demonstrates how to build a mostly autonomous HR assistant using ChatGPT (gpt-3.5-turbo) and LangChain's Agents and Tools modules. The system leverages chain-of-thought reasoning and the ReAct (Reasoning + Acting) loop to decompose complex HR queries into multi-step actions. It is equipped with three tools: Timekeeping Policies, Employee Data, and a Calculator, enabling it to answer timekeeping questions, retrieve employee information, and perform calculations. The prototype illustrates how LangChain agents can overcome LLM limitations—such as lack of access to proprietary or current data—by routing natural language queries to the appropriate external tool, and how process supervision (rewarding correct intermediate reasoning) reduces hallucinations and improves reliability.

Key Features

LangChain Agents with ReAct (Thought/Action/Action Input/Observation) loop for multi-step reasoning

Three customizable MRKL-based tools: Timekeeping Policies, Employee Data, Calculator

Chain-of-thought prompting to reduce hallucinations and improve reasoning

Router (LLM) that directs NL queries to the appropriate expert tool

Built on gpt-3.5-turbo (ChatGPT model) for conversational HR interactions

Prototype demonstrates access to proprietary or current data via external tool integration

Pros & Cons

Pros

Leverages chain-of-thought reasoning to enhance problem-solving and reduce hallucinations
Overcomes LLM limitations by routing queries to external tools for current or proprietary data
Modular design allows easy addition of new tools (e.g., more HR policies, APIs)
Demonstrates practical application of process supervision concepts from OpenAI/Google papers
Open-source prototype can be adapted and extended by developers

Cons

Only 'mostly autonomous' – still requires explicit tool configuration and fallback handling
Dependency on external LLM (gpt-3.5-turbo) and LangChain framework
May not handle highly nuanced or ambiguous HR queries without careful tool design
Prototype status – not tested at scale or for production security/compliance
Requires technical expertise to set up and maintain the tool integrations

Best For

Answering employee timekeeping policy questionsRetrieving specific employee data from a company databasePerforming calculations related to hours, payroll, or analyticsProviding a conversational interface for HR self-servicePrototyping enterprise applications that combine LLM reasoning with internal data sources

FAQ

What is the ReAct loop used in the HR assistant?

The ReAct (Reasoning + Acting) loop is a Thought/Action/Action Input/Observation cycle that combines chain-of-thought reasoning with acting. The LLM first generates a thought explaining its reasoning, then decides an action (e.g., calling a tool), provides the input, and observes the output before repeating until the final answer is reached. This approach, inspired by Google's chain-of-thought research, improves reasoning and reduces hallucinations.

How does the assistant access employee data without training on it?

The assistant uses LangChain's Tools module, which is based on the MRKL system. The LLM acts as a router: when a user asks for employee data, the LLM identifies the appropriate expert tool (e.g., Employee Data tool) and sends the query to that tool, which can access a proprietary database. The LLM never sees the raw data directly but uses the tool's output to form its response.