Antfly support-agent templates

Customer-ready templates for building documentation and knowledge support agents powered by Antfly retrieval.

The collection separates one portable support-agent contract from thin harness adapters. Customers can bring their own product documentation, Antfly table, model provider, branding, and support workflow without rewriting retrieval rules for every agent framework.

Why these templates are different

These are production-oriented retrieval templates, not merely examples of how to connect an MCP server to an agent:

• Adaptive retrieval uses semantic-first search for conceptual questions and hybrid BM25 plus vector RRF for exact technical questions.
• Answer quality and latency are evaluated together through bounded fallbacks, warm-session reuse, evidence checks, benchmarks, and regression tests.
• Chunk-level evidence, clean generation context, and citation deduplication keep answers grounded without discarding multiple useful chunks from a source.
• Read-only credentials, strict tool allowlists, bounded retries, and explicit failure behavior provide safe defaults.
• One retrieval contract works across Next.js, Codex, Claude, OpenAI Agents, Google ADK, Copilot, n8n, Hermes, and Pi.
• Operational metrics distinguish semantic-first and hybrid latency, fallback rate, payload size, source quality, and citation accuracy.

Available harnesses

• Knowledge Support — Next.js/Vercel product template
• Next.js and Vercel
• Claude Code
• Claude Agent SDK
• Hermes Agent
• Pi Agent Harness
• n8n
• Microsoft Copilot Studio
• Google ADK / Agents
• OpenAI Agents SDK
• Codex

Every harness will also include a working Antfly Docs Support example backed by the Antfly documentation table.

Shared contract

The shared directory defines behavior that every integration must preserve:

• grounded answers based on retrieved chunks
• one intent-selected query first, with at most one focused fallback
• semantic-first retrieval for broad concepts and hybrid BM25/vector RRF for exact terms
• read-only MCP access
• friendly citations instead of private object-storage paths
• deterministic escalation when evidence is insufficient
• common environment variables, evaluations, and failure handling

See the customer quickstart, template architecture, internal harness comparison, and validation gates.

Customer path

1. Load documentation into an Antfly document-search table.
2. Verify extraction, chunking, full-text indexing, and vector indexing.
3. Create an instance-scoped, read-only API key.
4. Pick a harness template.
5. Copy its example environment file and set the Antfly endpoint, key, table, and index names.
6. Customize product identity, documentation links, and escalation contact.
7. Run the shared smoke tests and evaluation set before publishing.

Working Antfly Docs examples

• The Knowledge Support template links the deployable antflydb/knowledge-support application with @antfly/components to provide the path from connected knowledge to hosted and embedded support.
• The Next.js/Vercel reference is a deployable application.
• The n8n recipe captures the working visual workflow and its bounded system message.
• The Copilot Studio recipe includes MCP setup, its 10/10 evaluation refinement, and Microsoft 365 publication.
• The Google ADK template connects Gemini-powered agents to Antfly MCP.
• The OpenAI Agents SDK template provides a deployable Python agent with Streamable HTTP MCP.
• The Codex template provides project MCP configuration and durable AGENTS.md instructions.
• Claude Code, Claude Agent SDK, Hermes, and Pi reuse the same table and query contract; their local configuration or code checks are documented in the validation matrix.

The example values target antfly_docs, document_vectors, and the text chunk field. Customers replace those values without changing the safety or retrieval behavior.