Building ReplyMe: Conversational AI & RAG Pipelines with Rasa

Customer support automation often struggles with intent matching, context retention, and domain-specific knowledge lookup. ReplyMe was architected to bridge this gap using Rasa Conversational AI combined with Retrieval-Augmented Generation (RAG) pipelines.

1. High-Level Architecture Overview

ReplyMe integrates traditional NLU intent classification with generative vector retrieval:

1. Rasa Dialogue Manager: Manages user session state, slot filling, and structured intent classification for transactional actions (e.g. refund requests, account status checks).
2. RAG Knowledge Base Pipeline: Intercepts open-ended queries, retrieves semantic context chunks from indexed documentation, and synthesizes accurate responses using LLMs.
3. Knowledge Indexer: Converts PDF/Markdown documentation into dense vector embeddings for sub-second similarity lookup.

2. RAG Retrieval & Context Scoring

pythonExample
# Python Context Retrieval Vector Pipeline
from langchain.vectorstores import Qdrant
from langchain.embeddings import OpenAIEmbeddings

def retrieve_knowledge_context(query: str, top_k: int = 3):
    embeddings = OpenAIEmbeddings()
    vector_db = Qdrant.connect(url="http://localhost:6333", collection_name="replyme-docs")
    docs = vector_db.similarity_search_with_score(query, k=top_k)
    
    return "\n---\n".join([doc[0].page_content for doc in docs if doc[1] > 0.78])

3. Key Achievements & Production Impact

• Accurate Intent Matching: Combined Rasa state management for predictable multi-turn dialogs with RAG fallback for open-ended queries.
• Support Workload Reduction: Handled over 70% of tier-1 support queries automatically without human agent intervention.
• Knowledge Base Indexing: Enabled instantaneous ingestion of updated company documentation within minutes.