Autonomous HR Chatbot

Implements a LangChain-based agent framework that interprets natural language HR queries and autonomously selects from three specialized tools (policy retrieval, employee data access, mathematical calculations) to compose answers. The agent uses chain-of-thought reasoning to decompose questions into tool invocations, managing context and tool outputs across multiple reasoning steps without human intervention.

Implements a RetrievalQA tool that converts HR policy documents into OpenAI text-embedding-ada-002 embeddings, stores them in Pinecone vector database, and retrieves semantically relevant policy excerpts at query time. The tool performs cosine similarity search to find policy sections matching the user's natural language question, enabling the agent to ground answers in actual HR documentation without hallucination.

Implements a PythonAstREPLTool that allows the agent to execute Python code against a pandas DataFrame containing employee records. The agent can generate and execute Python queries (e.g., 'df[df.name == "John"].salary') to access employee information, enabling dynamic data filtering without pre-defined query templates. The tool uses AST parsing to validate code safety before execution.

Implements an LLMMathChain tool that allows the agent to perform mathematical calculations (e.g., PTO accrual, salary adjustments, benefit deductions) by having the LLM generate Python math expressions and executing them. The tool handles unit conversions and multi-step calculations, enabling the agent to answer HR questions requiring numerical reasoning without hardcoding calculation logic.

Implements a Streamlit frontend (hr_agent_frontend.py) that renders a chat interface using the streamlit_chat component, allowing users to submit HR queries and view agent responses in a familiar conversation format. The frontend manages session state to maintain conversation history and handles streaming responses from the backend, providing real-time feedback to users.

Implements two backend modules (hr_agent_backend_local.py and hr_agent_backend_azure.py) that abstract the LLM provider and deployment environment, allowing the same agent logic to run against OpenAI API (local) or Azure OpenAI Service (cloud). Both backends use the same LangChain agent interface, enabling seamless switching between deployment targets without code changes to the agent logic.

Implements a Jupyter notebook (store_embeddings_in_pinecone.ipynb) that processes HR policy documents through a multi-step pipeline: splitting documents into semantic chunks, generating embeddings using OpenAI's text-embedding-ada-002 model, and storing embeddings in Pinecone with metadata. This pipeline runs offline before the agent starts, enabling fast semantic search at query time without embedding computation overhead.

Implements employee data management through a CSV file that is loaded into a pandas DataFrame at agent startup. The system stores employee records with fields like name, department, salary, and hire_date, making employee data accessible to the PythonAstREPLTool for dynamic querying. This approach avoids database dependencies while supporting basic employee data operations.

Enables the agent to decompose complex HR questions into multiple sequential steps, invoking different tools (policy retrieval, employee data queries, calculations) in a chain to compose final answers. For example, a query like 'What is the PTO balance for John Smith?' is decomposed into: (1) retrieve PTO policy, (2) query John's hire date, (3) calculate accrual based on tenure. The agent manages intermediate results and passes them between tool invocations.

Manages conversation history in the Streamlit frontend, maintaining a record of user queries and agent responses across multiple turns. The agent uses this history to provide context for follow-up questions, enabling multi-turn conversations where the agent can reference previous answers. Session state is stored in Streamlit's session_state object, allowing history to persist within a single user session.