feat: implement CLI entry point for support ticket triage

.env.example

@@ -0,0 +1,5 @@
+LLM_PROVIDER=auto
+ANTHROPIC_API_KEY=
+ANTHROPIC_MODEL=claude-sonnet-4-6
+OPENAI_API_KEY=
+OPENAI_MODEL=gpt-4o-mini

.gitignore

@@ -11,6 +11,7 @@ node_modules/
 venv/
 __pycache__/
 *.pyc
+.cache/
 .pytest_cache/
 .mypy_cache/
 .ruff_cache/

code/README.md

@@ -0,0 +1,85 @@
+# Support Triage Agent — HackerRank Orchestrate (May 2026)
+
+Terminal-based RAG agent that triages support tickets across HackerRank, Claude, and Visa using only the local corpus in `../data/`.
+
+## Architecture
+
+```
+ticket ──► pre-rules (regex)           ──► early return (escalate / out-of-scope)
+       ──► company inference (None→best by retrieval)
+       ──► language gate (non-English → escalate for None/Visa)
+       ──► dense retrieval (MiniLM, single index, metadata filter)
+       ──► coverage-floor check         ──► early escalate
+       ──► LLM (Anthropic primary, OpenAI fallback, tool-forced JSON)
+       ──► extractive fallback when no LLM key/quota is available
+       ──► verifier (citation + sentence grounding)
+       ──► post-rules (confidence/citation/area allow-list)
+       ──► RowOutput
+```
+
+## Setup
+
+```bash
+cd code/
+python -m venv .venv && source .venv/bin/activate
+pip install -r requirements.txt
+cp .env.example .env  # set ANTHROPIC_API_KEY and/or OPENAI_API_KEY
+```
+
+Recommended model: `OPENAI_MODEL=gpt-4o` (tuned against this) or `ANTHROPIC_MODEL=claude-sonnet-4-6`. `gpt-4o-mini` over-escalates sensitive-but-answerable tickets.
+
+## Run
+
+```bash
+# Full pipeline against the real tickets:
+python main.py
+
+# Eval against the 10-row gold sample:
+python eval.py
+
+# Resume after partial run:
+python main.py --resume
+
+# No embeddings (TF-IDF fallback, fully offline):
+python main.py --no-embeddings
+
+# Dry run (skip LLM, only pre-rule decisions):
+python main.py --dry-run --limit 5
+```
+
+Reads input from `../support_tickets/support_tickets.csv` and writes to
+`../support_tickets/output.csv` with the exact 8-column header
+`issue,subject,company,response,product_area,status,request_type,justification`.
+
+## Design decisions
+
+- **Single dense index, not BM25 + RRF.** Corpus is small (~3k chunks); MiniLM cosine is enough and avoids index-mismatch bugs under time pressure.
+- **Rules before LLM, rules after LLM.** Pre-rules catch sensitive cases (fraud, legal, refund demands, score appeals, prompt injection, outage). Post-rules enforce confidence floor, mandatory citations, and product_area allow-list. The LLM is only trusted on grounded support questions.
+- **Coverage floor check.** If retrieval similarity is below threshold we escalate before calling the LLM — this is the main hallucination defense.
+- **Grounding verifier.** Cited chunk IDs must come from retrieval; unsupported response sentences are stripped or the row is escalated.
+- **Determinism.** `temperature=0`, fixed `seed=42`, sorted tie-breaks, stable cosine sort, deterministic chunking.
+- **Secrets.** Read only from env (`ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, model/provider settings). Local `.env`, `.env.local`, `code/.env`, and `code/.env.local` are supported.
+- **`None`-company area handling.** When `company=None` and a pre-rule short-circuits the row, `product_area` is left blank for pleasantries and pure escalations, but `off_topic_trivia` is tagged `conversation_management` (Claude's catch-all for off-topic chat) since None-company invalid questions belong to that bucket.
+
+## Files
+
+| File | Purpose |
+|---|---|
+| `main.py` | CLI entry point |
+| `agent.py` | Orchestrator |
+| `retriever.py` | Dense retriever + TF-IDF fallback |
+| `corpus.py` | Markdown loader + chunker |
+| `escalation.py` | Pre/post rule tables |
+| `verifier.py` | Citation grounding check |
+| `prompts.py` | System prompt + few-shots |
+| `llm_client.py` | Anthropic/OpenAI tool-call wrapper |
+| `schemas.py` | Pydantic models |
+| `io_csv.py` | CSV reader/writer with strict header |
+| `eval.py` | Accuracy harness vs. gold sample |
+| `config.py` | Constants |
+
+## Known limits
+
+- Gold sample is only 10 rows — accuracy numbers are noisy.
+- Visa corpus is 14 docs; Visa tickets escalate often by design.
+- We do not split multi-intent tickets; the LLM handles them in a single call.