LangChain RAG onboarding assistant

Company policy answers with cited sources.

Inductify lets new hires query 10K+ internal policy documents and receive grounded answers with source citation cards. I designed the retrieval and generation pipeline around fast Q&A, document indexing, and evaluation-driven improvements to answer quality.

  • 10K+ indexed internal docs
  • <5s cited Q&A latency target
  • 83% held-out RAG eval accuracy
View screenshots See architecture
Inductify assistant answering a PTO rollover question with cited source excerpts

Product Screens

The interface is built around a familiar chat workspace: conversation history, source-backed answers, document upload, dark mode, and clear onboarding prompts.

LangChain RAG ChromaDB Python FastAPI OpenAI LLM CrossEncoder Docker
Inductify empty chat workspace with onboarding assistant prompt

Onboarding Workspace

A chat-first interface gives employees a direct place to ask questions about policies, benefits, and procedures.

Inductify PTO answer with a cited source card

Source-Grounded Answers

Answers include retrieved source snippets and confidence scores, making policy responses easier to verify.

Inductify benefits answer with cited policy source

Benefits and HR Queries

The assistant handles common onboarding questions such as PTO, health benefits, remote work, and reimbursements.

Inductify profile menu and remote work answer

Conversation Context

Recent conversations and profile actions keep the workflow familiar while preserving multi-turn onboarding sessions.

Inductify dark mode answer with source cards

Dark Mode

The UI supports light and dark themes, keeping the assistant comfortable for long policy-review sessions.

Inductify dark mode travel expense answer

Policy Detail Retrieval

Retrieved context helps the assistant answer detailed questions about limits, eligibility, timelines, and required receipts.

Architecture

Inductify separates the chat UI from the retrieval and generation layer. A modular FastAPI backend exposes RESTful endpoints for real-time Q&A and asynchronous document indexing, while the RAG pipeline uses OpenAI embeddings, ChromaDB vector search, and an optional CrossEncoder reranker before generation.

1

Document Intake

Employees upload txt, md, PDF, or xlsx files for indexing through the chat interface.

2

Indexing Pipeline

FastAPI chunks documents, creates OpenAI embeddings, and stores vectors in ChromaDB.

3

RAG + Reranking

LangChain retrieves top-k chunks and a CrossEncoder reranker improves source relevance before generation.

4

Evaluation Loop

Offline tests track top-3 accuracy and compare LLM-only answers against the full RAG pipeline.