SupportRAG CLI

Turn any public support/help-center page into a local support-ticket triage agent.

SupportRAG is a deterministic CLI product for support teams, founders, and hackathon builders. It crawls support pages into a local Markdown corpus, retrieves relevant docs with BM25, applies safety/escalation rules, and outputs a clean support triage CSV.

Why it is useful

Most support bots hallucinate. SupportRAG is designed for high-trust support workflows:

• Local corpus first. No unsupported web guessing.
• Deterministic BM25 retrieval.
• Rule-based escalation for risky cases.
• Optional LLM paraphrasing seam, disabled by default.
• CSV in, CSV out. Easy to plug into ops workflows.
• Works as a hackathon submission and a future product foundation.

Install locally

git clone https://github.com/princepal9120/hackerrank-orchestrate-may26.git
cd hackerrank-orchestrate-may26
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .

This installs two equivalent commands:

supportrag --help
supportbot --help

Without installation, use the repo-local launcher:

```bash
python3 supportrag.py --help

## Quickstart with any support page

### 1. Build a local corpus

```bash
python3 supportrag.py ingest \
  --url https://example.com/support \
  --output corpora/example-support \
  --max-pages 100

2. Create tickets CSV

Issue,Subject,Company
"How can I reset billing?","Billing help","ExampleCo"
"The whole product is down","Urgent","ExampleCo"

3. Triage tickets

python3 supportrag.py triage \
  --data corpora/example-support \
  --input tickets.csv \
  --output output.csv

Output columns:

issue,subject,company,response,product_area,status,request_type,justification

Allowed values:

• status: replied, escalated
• request_type: product_issue, feature_request, bug, invalid

HackerRank Orchestrate submission mode

The original challenge entrypoint still works:

python3 code/main.py \
  --data data \
  --input support_tickets/support_tickets.csv \
  --output support_tickets/output.csv

Product architecture

URL support page
  -> ingest_url.py
  -> local Markdown corpus
  -> corpus.py chunking
  -> retrieval.py BM25
  -> rules.py safety router
  -> agent.py orchestration
  -> validation.py schema guard
  -> output.csv

What makes this product-ready

• Installable Python package via pyproject.toml.
• Real CLI commands: ingest, triage, eval.
• No required third-party dependencies.
• Local-first support corpus design.
• Clear upgrade path to SaaS:
  • corpus registry
  • incremental recrawls
  • embedding reranker
  • tenant-specific safety policies
  • dashboard/API
  • citations and confidence scores

Current limitations

• The crawler is a deterministic stdlib HTML crawler, not a browser. JavaScript-heavy docs may need a future Playwright ingestion backend.
• Safety rules are currently tuned for the HackerRank/Claude/Visa challenge and should become configurable policies for production customers.
• Optional LLM paraphrasing is intentionally disabled by default for reproducibility.