🧠 Conversational Insights

AI-powered talking point generation for client advisors — built on Multi-Agent Systems (MAS), Agentic RAG, and intelligent information triaging.

---

Overview

Conversational Insights is a LangGraph / LangChain-powered backend that helps private banking client advisors prepare for client meetings by automatically surfacing relevant talking points from past interactions.

The system ingests emails and call transcripts, stores them in a vector database, and uses an agentic retrieval pipeline to answer advisor queries grounded in historical context.

---

Key Features

Feature	Description
📥 ETL Pipeline	Ingests email PDFs, extracts structured metadata via LLM, and stores chunked vectors in ChromaDB
🤖 Agentic RAG	A LangGraph ReAct agent that retrieves relevant context before answering — never hallucinates without evidence
🏛️ Multi-Agent System (MAS)	Modular agent architecture designed for extension with specialised sub-agents (e.g. triage, summarisation, compliance)
🗂️ Information Triaging	Retrieval is semantically ranked so the most relevant past interactions are surfaced first
🔒 Prompt Injection Safe	Retrieved content is treated as data only; embedded instructions are ignored

---

Architecture

┌─────────────────────────────────────────────┐
│                  main.py                     │
│           (advisor query entrypoint)         │
└────────────────────┬────────────────────────┘
                     │
          ┌──────────▼──────────┐
          │   Retrieval Agent   │  ← LangGraph ReAct Agent
          │  (retrieval_agent)  │
          └──────────┬──────────┘
                     │  tool call
          ┌──────────▼──────────┐
          │  retrieve_context   │  ← similarity search (k=2)
          └──────────┬──────────┘
                     │
          ┌──────────▼──────────┐
          │      ChromaDB       │  ← populated by ETL pipeline
          │   (email vectors)   │
          └─────────────────────┘

ETL Pipeline (run separately):
  data/emails/*.pdf  →  PyPDFLoader  →  LLM metadata extraction
  →  RecursiveCharacterTextSplitter  →  OpenAI Embeddings  →  ChromaDB

---

Project Structure

conversational-insights/
├── main.py                        # Query entrypoint
├── script/
│   └── run_etl_pipeline.py        # Run email ingestion
├── src/
│   ├── config.py                  # Central config (models, paths, keys)
│   ├── agents/
│   │   └── retrieval_agent.py     # LangGraph ReAct retrieval agent
│   ├── etl/
│   │   └── etl_pipeline.py        # Email PDF → ChromaDB pipeline
│   ├── models/
│   │   └── models.py              # Pydantic schemas (EmailMetadata)
│   ├── prompts/
│   │   └── retrieve_agent_prompt.py
│   ├── tools/
│   │   └── retrieve_content.py    # LangChain retrieval tool
│   └── utils/
│       └── vector_store.py        # Shared ChromaDB instance
└── data/
    ├── emails/                    # Raw email PDFs (input)
    ├── transcripts/               # Call transcripts
    └── vector/                    # ChromaDB persist directory

---

Quickstart

1. Prerequisites

• Python ≥ 3.14
• uv package manager
• OpenAI API key

2. Install dependencies

uv sync

3. Configure environment

cp .env.example .env
# Fill in your OPENAI_API_KEY

4. Run the ETL pipeline

Place your email PDFs in data/emails/, then:

uv run python3 ./script/run_etl_pipeline.py

This will extract metadata, chunk each email, embed it, and persist to data/vector/.

5. Query the agent

uv run python3 main.py

Edit the query variable in main.py to ask any question about client interactions.

---

Configuration

All tunable constants live in src/config.py:

Variable	Default	Description
CHAT_MODEL	gpt-4o-mini	LLM used for metadata extraction and the agent
EMBEDDING_MODEL	text-embedding-3-small	OpenAI embedding model
CHROMA_COLLECTION_NAME	email_collection	ChromaDB collection
CHROMA_PERSIST_DIRECTORY	data/vector/	Where vectors are persisted
CHUNK_SIZE	512	Token chunk size for text splitting
CHUNK_OVERLAP	100	Overlap between chunks

---

🚧 Work In Progress

Transcript & Message Analysis

Call transcripts are already present in data/transcripts/ but are not yet ingested or searchable. Planned work:

• Transcript ETL — a parallel ingestion pipeline for .txt call transcripts (alongside the existing email PDF pipeline), with its own metadata schema (caller, client, call type, date)
• Unified retrieval — a single vector collection (or separate namespaced collections) covering both emails and transcripts, so the agent can reason across all interaction types
• Interaction summarisation — an LLM-powered summarisation step that distills long transcripts into key points before embedding, improving retrieval quality

Multi-Agent System (MAS) for Query Handling

The current system is a single retrieval agent. The planned MAS will introduce a supervisor + specialist architecture:

User Query
    │
    ▼
┌─────────────────────┐
│   Supervisor Agent  │  ← routes query to the right specialist(s)
└──────┬──────┬───────┘
       │      │
  ┌────▼──┐ ┌─▼──────────────┐
  │Triage │ │  Retrieval      │
  │Agent  │ │  Agent(s)       │  ← email / transcript specialists
  └───────┘ └─────────────────┘
       │
  ┌────▼────────────────┐
  │ Synthesis Agent     │  ← aggregates results → talking points
  └─────────────────────┘

• Triage Agent — classifies the query type (investment discussion, trade instruction, payment request, general enquiry) and decides which specialists to invoke
• Specialist Retrieval Agents — domain-scoped agents for emails and transcripts respectively
• Synthesis Agent — aggregates retrieved context and generates structured talking points for the advisor

---

Tech Stack

• LangChain — tool definitions, prompt construction, LLM wrappers
• LangGraph — stateful agent graph (ReAct loop)
• ChromaDB — local vector store
• OpenAI — gpt-4o-mini + text-embedding-3-small
• Pydantic — structured LLM output schemas
• uv — dependency and environment management