Network Insights RAG for Cisco Catalyst Center

NetworkInsights is a Retrieval-Augmented Generation (RAG) toolkit for network operations. It collects CLI outputs from Cisco Catalyst Center, builds embeddings in Chroma, and provides interactive query/conversation clients for OpenAI and Anthropic models. Written by Gabi Zapodeanu, Principal TME, updated with Cursor/ Codex 5.3.

Repository Structure

• DB_Server/
  • chroma_server.py - starts the Chroma server.
  • chroma_create_erase_collection.py - create/delete collection utility.
• Data_Collection/
  • network_data_collection.py - collects CLI outputs from listed devices.
  • device_list.yaml - devices to query.
  • cli_commands.yaml - commands to execute.
• Transform_Data/
  • embeddings_to_chroma.py - splits documents, adds metadata, uploads embeddings.
• Client_Apps/
  • Classic clients:
    • query_network_insights_openai.py
    • conversation_network_insights_openai.py
    • conversation_network_insights_anthropic.py
  • LCEL clients:
    • query_network_insights_openai_lcel.py
    • conversation_network_insights_openai_lcel.py
    • conversation_network_insights_anthropic_lcel.py
  • Test helper:
    • test_lcel_versions.py
• DATASET/ - raw text data used for embeddings.
• VDB/ - local Chroma vector database storage.
• environment.env - local runtime configuration (do not commit secrets).

Features

• End-to-end network data flow:
  1. Collect device CLI data from Catalyst Center
  2. Build metadata-rich embeddings
  3. Query with RAG clients
• Metadata-aware retrieval (minimal filter):
  • device filter based on explicit hostname in query
• Model/provider flexibility:
  • OpenAI and Anthropic clients
  • Classic and LCEL variants
• Enterprise-friendly HF TLS handling:
  • supports custom CA bundle via env configuration

Requirements

• Python 3.10+
• Cisco Catalyst Center access (for data collection)
• OpenAI and/or Anthropic API credentials (for client apps)

Setup

1. Create and activate a virtual environment:

python3 -m venv .venv
source .venv/bin/activate

2. Install dependencies:

pip install -r requirements.txt

3. Configure environment.env (local only):

• API keys
• Catalyst Center credentials
• Chroma settings
• model settings

Recommended keys:

OPENAI_MODEL='gpt-5.5'
CLAUDE_MODEL='claude-sonnet-4-6'
MODEL_NAME='all-MiniLM-L6-v2'

Optional keys:

MODEL_LOCAL_PATH='/absolute/path/to/local/embedding/model'
HF_CA_BUNDLE='/absolute/path/to/ca-bundle.pem'

Typical Workflow

1) Start Chroma

.venv/bin/python DB_Server/chroma_server.py

2) Create collection (if needed)

.venv/bin/python DB_Server/chroma_create_erase_collection.py

3) Collect network data

Update:

• Data_Collection/device_list.yaml
• Data_Collection/cli_commands.yaml

Run:

.venv/bin/python Data_Collection/network_data_collection.py

4) Build embeddings

.venv/bin/python Transform_Data/embeddings_to_chroma.py

5) Run clients

Classic query:

.venv/bin/python Client_Apps/query_network_insights_openai.py

Classic conversations:

.venv/bin/python Client_Apps/conversation_network_insights_openai.py
.venv/bin/python Client_Apps/conversation_network_insights_anthropic.py

LCEL clients:

.venv/bin/python Client_Apps/query_network_insights_openai_lcel.py
.venv/bin/python Client_Apps/conversation_network_insights_openai_lcel.py
.venv/bin/python Client_Apps/conversation_network_insights_anthropic_lcel.py

LCEL Smoke Test

Run all LCEL clients with predefined prompts:

.venv/bin/python Client_Apps/test_lcel_versions.py

Notes

• environment.env is for local development and should not contain shared secrets.
• HF warning about unauthenticated requests is non-blocking for most POC usage.
• For best retrieval precision, include device hostname explicitly in queries.