olasunkanmi-SE/deriv

Perfil de olasunkanmi-SE

Última atividade de origem registrada: 17 de maio de 2026 às 11:18

Catálogo do SkillsMP atualizado: 26 de outubro de 2026 às 11:37

skills coletadas: 3

Estrelas no GitHub: 1
Forks no GitHub: 0

README do repositório

Arquivo de origem do README: README.md

Ler o README original

Expandir README completo

# Deriv AI Support Triage Pipeline An AI-assisted pipeline that ingests customer support tickets and a local knowledge base, then runs four sequential LLM stages to classify, prioritise, draft responses for, and validate each ticket — producing a full set of auditable JSON artifacts. ## Prerequisites - **Node.js 20+** - **Anthropic API key** — set as `ANTHROPIC_API_KEY` in your environment ## Quick start ```bash npm install export ANTHROPIC_API_KEY=sk-ant-... npm start ``` To validate the produced artifacts against all pipeline rules: ```bash npm run validate ``` ## npm scripts | Command | What it does | |---|---| | `npm start` | Run the full pipeline (`tickets.json` + `knowledge_base/` → `artifacts/`) | | `npm run validate` | Check all 10 output artifacts exist, are valid JSON, and pass schema/vocab rules | | `npm test` | Run all tests (unit + integration) with Vitest | | `npm run test:unit` | Unit tests only | | `npm run test:integration` | Integration test (full pipeline with stub LLM, no API key needed) | | `npm run test:watch` | Vitest watch mode | | `npm run build` | Compile TypeScript to `dist/` | | `npm run typecheck` | Type-check without emitting | ## Pipeline stages The pipeline advances through 13 explicit states. Each stage asserts its predecessor before running. ``` INIT → INPUTS_LOADED → KNOWLEDGE_INDEXED → TICKETS_NORMALISED → RETRIEVAL_COMPLETE → TRIAGE_COMPLETE (LLM stage 1) → RESPONSE_DRAFTED (LLM stage 2) → ACTION_PLAN_CREATED (LLM stage 3) → GROUNDING_VALIDATED (LLM stage 4 + deterministic checks) → FINAL_OUTPUTS_WRITTEN → QUEUE_RANKED → AUDIT_LOG_EXPORTED ``` | Stage | Artifact produced | Model | |---|---|---| | BM25 retrieval | `artifacts/retrieval_results.json` | — | | Triage | `artifacts/triage.json` | claude-haiku-4-5 | | Response drafting | `artifacts/response_drafts.json` | claude-sonnet-4-6 | | Action planning | `artifacts/action_plan.json` | claude-sonnet-4-6 | | Grounding validation | `artifacts/grounding_validation.json` | claude-sonnet-4-6 | | Finalisation | `artifacts/final_ticket_outputs.json` | — | | Queue ranking | `artifacts/queue_ranking.json` | — | | Audit log | `artifacts/audit_log.json` | — | Every LLM call is appended to `artifacts/llm_calls.jsonl` with stage, model, prompt hash, and ticket/chunk IDs. ## Sample output **`artifacts/triage.json`** (excerpt): ```json [ { "ticket_id": "T-1002", "category": "account_security", "urgency": "critical", "sentiment": "negative", "resolution_mode": "needs_specialist_escalation", "recommended_queue": "trust_and_safety", "reasoning_summary": "VIP customer has locked account with open positions at risk during security review..." } ] ``` **`artifacts/queue_ranking.json`** (excerpt): ```json [ { "ticket_id": "T-1002", "rank": 1, "score": 59, "urgency": "critical", "customer_tier": "vip" }, { "ticket_id": "T-1004", "rank": 2, "score": 48, "urgency": "critical", "customer_tier": "standard" } ] ``` Scoring: urgency (critical=40, high=30, medium=20, low=10) + customer tier (vip=+10) + category priority (account_security=+9, withdrawal_delay=+10, etc.). ## Providing inputs **Tickets** — edit `tickets.json` at the repo root. Each ticket requires: ```json { "ticket_id": "T-XXXX", "submitted_at": "2026-05-10T09:15:00Z", "customer_tier": "standard", "language": "en", "subject": "...", "message": "..." } ``` `customer_tier` accepts `standard` or `vip`. **Knowledge base** — add or edit Markdown files under `knowledge_base/`. The chunker splits on `#`/`##` headings; each chunk gets a stable ID (`{document_id}-{NNN}`) and a SHA-256 content hash. The BM25 retriever selects the top-4 most relevant chunks per ticket. ## Environment variables | Variable | Default | Description | |---|---|---| | `ANTHROPIC_API_KEY` | — | Required; pipeline exits with a clear error if absent | | `TRIAGE_MODEL` | `claude-haiku-4-5-20251001` | Model for triage stage | | `RESPONSE_MODEL` | `claude-sonnet-4-6` | Model for response drafting, action planning, grounding | | `RETRIEVAL_TOP_K` | `4` | Chunks retrieved per ticket | | `RETRIEVAL_MIN_SCORE` | `0.1` | BM25 min score; tickets below threshold are flagged `low_retrieval_confidence` | ## Running tests without an API key The integration test runs the full pipeline end-to-end using a `StubLLMService` that returns fixture responses — no API key required: ```bash npm run test:integration ``` Unit tests are also entirely offline: ```bash npm run test:unit ``` ## Architecture The codebase follows Clean Architecture with a strict one-way import rule: ``` domain ← application ← infrastructure ← interface ``` - **`src/domain/`** — entities, value objects, repository/service interfaces, `PipelineState` enum. No dependencies. - **`src/application/`** — use cases, prompt builders, `PipelineOrchestrator`, `GroundingChecker`. Depends only on domain. - **`src/infrastructure/`** — `AnthropicLLMService`, `TfIdfBm25Retriever`, `MarkdownChunker`, file repositories, `JsonlLLMCallLogger`. - **`src/interface/`** — `cli/run.ts` (DI wiring + pipeline entry), `validation/validate.ts`. See `architecture.md` for the full design document and `assumptions.md` for the retrieval confidence and fallback policy.

Carregando o README do repositório…

Skills neste repositório

Desenvolvedores de software Analistas de garantia de qualidade de software e testadores

2 categorias ocupacionais · 100% classificado

skill

ocupação

descrição

atualizado

optimize-code

Desenvolvedores de software

Transforms sub-optimal code into a high-performance version with measurable algorithmic and resource improvements.

Idioma do texto original: inglês

2026-05-17

refactor-code

Desenvolvedores de software

Transforms complex code into clean, highly maintainable code following SOLID principles and design patterns.

Idioma do texto original: inglês

2026-05-17

review-code

Analistas de garantia de qualidade de software e testadores

Performs a compact, elite-level structured code review using the CodeBuddy framework.

Idioma do texto original: inglês

2026-05-17

Mostrando 3 de 3 skills coletadas.

Instalar com um assistente de IA

Copie este prompt para o assistente de IA que você usa.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/olasunkanmi-SE/deriv.