An assistant grounded in your documentation answers customers and agents with citations, files the ticket, and escalates with context, without a single conversation leaving your infrastructure.
Every support team evaluates AI. Most evaluations end on the same four findings, and none of them are about model quality.
The pipeline that turns a documentation folder into a support channel people can actually trust.
02
Hybrid retrieval plus reranking; every answer cites the article it came from.
04
Unresolved cases hand off to a person with the full context attached.
The scenario decides the form: a backend the helpdesk points at, or support intelligence inside the product you ship.
Serve it
Server-side agents bundle prompt, knowledge, tools, and memory; the helpdesk, the chat widget, and internal tools adopt them by name over the APIs they already speak, with identities, audit, and policies handled once.
Embed it
ISVs ship in-app assistants that answer from the product's own documentation, run offline at customer sites, and never add a cloud dependency to the deal.
The properties that separate a support assistant from a liability generator.
Grounded
Every answer cites the document and passage it came from, so agents verify in one click and customers get policy, not improvisation.
Governed
Tools carry risk metadata and run under allow, deny, and approval rules; the assistant files tickets because policy permits it, not because nothing stopped it.
Unmetered
No per-token fees means no rationing during incidents, launches, and renewals, when volume peaks.
The server agent is the high-level path: defined once, adopted by name. The OpenAI-compatible endpoint keeps existing integrations working, and the embedded form runs the same stack in process.
The high-level API: the agent bundles prompt, knowledge, tools, and memory on the server, and every channel adopts it with one field.
curl http://your-server:5189/lmkit/v1/chat \
-H "Authorization: Bearer your-server-token" \
-H "Content-Type: application/json" \
-d '{"agent":"support-assistant",
"messages":[{"role":"user","content":"Where do I renew my enterprise license?"}]}'
# The agent supplies system prompt, skill, tools, and memory;
# change the agent on the server and every channel follows.The OpenAI-compatible endpoint keeps the integrations your helpdesk already has working unchanged.
curl http://your-server:5189/v1/chat/completions \
-H "Authorization: Bearer your-server-token" \
-H "Content-Type: application/json" \
-d '{"model":"qwen3.5:9b",
"messages":[{"role":"user","content":"Where do I renew my enterprise license?"}]}'Ground a chat on the knowledge base and tune retrieval in two lines.
using LMKit.Retrieval;
var ragEngine = new RagEngine(embeddingModel);
ragEngine.ImportText(helpArticle,
new TextChunking() { MaxChunkSize = 400 },
dataSourceIdentifier: "help-center",
sectionIdentifier: "billing");
using var chat = new RagChat(ragEngine, chatModel);
chat.MaxRetrievedPartitions = 3;
chat.MinRelevanceScore = 0.5f;
var answer = await chat.SubmitAsync("Where do I renew my enterprise license?");The assistant composes documented capabilities you can inspect on their own.
Ground
Conversational retrieval with citations, four query modes, streaming.
How RAG chat worksAct
Ticketing, CRM, and status tools defined once, called under policy.
How tools workGovern
Allow, deny, and approval rules over every tool the assistant can reach.
How governance worksRemember
Customer context persists across sessions, retrieval-grounded and inspectable.
How memory worksRespond
Tokens stream to the chat widget as they decode; no dead air.
How streaming worksRelated
The general assistant pattern this use case specializes for support.
The assistant use caseGround it. Retrieval-augmented chat answers from your documentation and cites the document and passage behind every claim, and retrieval thresholds keep it from answering at all when nothing relevant exists, which is the honest failure mode.
Yes. LM-Kit One speaks OpenAI, Anthropic, and Ollama dialects plus MCP, so integrations built for those APIs point at your server by changing the base URL. Server-side agents bundle the prompt, knowledge, and tools so every channel gets the same behavior.
Permission policies. Every tool carries risk metadata and runs under allow, deny, and approval rules; sensitive actions can require human sign-off, and the audit trail records what was called and why.
It escalates with context: the conversation, the retrieved sources, and the attempted steps travel with the handoff, so the human agent starts informed instead of starting over.
Private AI for support