SAP AI SDK Support Bot

Automatically answers GitHub issues on SAP/ai-sdk-js using the SAP AI SDK itself — dogfooding @sap-ai-sdk/langchain to support @sap-ai-sdk.

Trigger: issues: opened → GitHub Actions (contributors only)
LLM: anthropic--claude-4.5-haiku via OrchestrationClient on SAP AI Core
Retrieval: Context7 MCP (docs) + GitHub MCP (issues) + local semantic search (embeddings)

---

How it works

GitHub issue opened (MEMBER / OWNER / COLLABORATOR only)
  │
  ├── ① Post issue preview to Slack          (continue-on-error)
  ├── ② Generate answer — reply.ts → askBot()
  │       └── Agent loop (max 8 iterations)
  │             LLM decides which tools to call:
  │             • context7__query-docs    — official SAP AI SDK docs (532 snippets)
  │             • github__search_issues   — past issues keyword search
  │             • github__get_issue       — full issue body
  │             • github__search_code     — code examples in the repo
  │             • semantic_search         — local embedding index (cross-validation)
  ├── ③ Reply bot answer in Slack thread     (continue-on-error)
  └── ④ Post comment on GitHub issue         (always runs)

---

Local setup

Prerequisites

• Node.js 22+, pnpm 10+
• SAP AI Core service key with anthropic--claude-4.5-haiku deployment
• GitHub Personal Access Token (repo read scope)

Install

pnpm install
cp .env.example .env
# Fill in AICORE_SERVICE_KEY and GITHUB_TOKEN in .env

Run

pnpm start    # interactive CLI (readline REPL)
pnpm serve    # HTTP server on :3000 (POST /ask, GET /health)
pnpm build    # TypeScript type check
pnpm refresh  # Rebuild local embedding index from GitHub issues

---

Environment variables

Variable	Required	Description
AICORE_SERVICE_KEY	✅	SAP AI Core service key (full JSON string)
GITHUB_TOKEN	✅	GitHub PAT — repo read scope
BOT_API_KEY	optional	Bearer token for POST /ask (skip for local dev)
PORT	optional	HTTP server port (default: 3000)
SLACK_BOT_TOKEN	optional	Slack Bot OAuth token for issue notifications
SLACK_CHANNEL_ID	optional	Slack channel ID (e.g. C...)

---

GitHub Actions setup

Secrets required (SAP/ai-sdk-js → Settings → Secrets → Actions)

Secret	Value
AICORE_SERVICE_KEY	SAP AI Core service key JSON
GITHUB_TOKEN	Auto-provided by GitHub Actions
SLACK_BOT_TOKEN	xoxb-... Bot token (optional)
SLACK_CHANNEL_ID	Slack channel ID (optional)

Workflows

File	Trigger	Purpose
.github/workflows/issue-reply.yml	issues: opened	Generate and post answer
.github/workflows/refresh-embeddings.yml	Weekly (Mon 03:00 UTC) + manual	Rebuild embedding index

Slack setup (optional)

1. api.slack.com/apps → Create New App → From scratch
2. OAuth & Permissions → Bot Token Scopes: chat:write, chat:write.public
3. Install to Workspace → copy Bot User OAuth Token (xoxb-...)
4. Add SLACK_BOT_TOKEN and SLACK_CHANNEL_ID secrets to the repo

---

Embedding index

The local semantic search uses pre-built embeddings (text-embedding-3-large, 3072 dimensions).

File	Contents	Count
src/embeddings-resolved.json	Closed issues with resolution	71
src/embeddings-open.json	Open issues	13

Rebuild with pnpm refresh or trigger refresh-embeddings.yml manually.

---

HTTP API

POST /ask
  Body: { "question": string }   (max 2000 chars)
  Auth: Authorization: Bearer <BOT_API_KEY>  (required if BOT_API_KEY is set)
  Response: { "answer": string }

GET /health
  Response: { "status": "ok" }

---

Project structure

src/
  agent.ts              — askBot(), initAgent(), closeAgent() — core agent loop
  reply.ts              — GitHub Actions entry point (parses issue body → askBot)
  server.ts             — HTTP server entry point
  cli.ts                — Local REPL entry point
  embeddings.ts         — loadEmbeddingIndex(), semanticSearch()
  knowledge.ts          — SDK_KNOWLEDGE static reference (injected into system prompt)
  embeddings-resolved.json
  embeddings-open.json
scripts/
  refresh-embeddings.ts — Rebuild embedding index from GitHub API
.github/workflows/
  issue-reply.yml
  refresh-embeddings.yml

---

Backlog

See IMPLEMENTATION_PLAN.md for planned improvements.