Key facts
| Agent runtime | Function calling loop on /v1/chat/completions (live) |
| Tools | Knowledge search, article fetch, freshness check and escalation |
| Models | 30+ models, route lookups and synthesis independently |
| Answer format | JSON mode for answer, sources and confidence (live) |
| Freshness | Updated-date metadata used to rank and warn on stale content |
| Audit | Per-turn tool calls and inference metadata exportable to SIEM |
| Deployment | Region-locked planes, VPC, on-prem and air-gapped options |
| Roadmap endpoints | Assistants-style managed endpoints are coming soon; the tool loop is live now |
TL;DR
- Specialise the agent for answering: retrieve, verify freshness, cite or escalate.
- Rank by authority and recency, not similarity alone.
- Return confidence with every answer, and route low-confidence cases to humans.
- Capture unanswered questions as a knowledge-gap backlog for content owners.
- Measure deflection and gap closure together — both must move.
How it works, step by step
- Define approved sources and their owners, and mark which sources outrank others when they disagree.
- Expose knowledge search, article fetch and escalation as tools with typed schemas and server-side permission checks.
- Implement the retrieval loop: rewrite the query, search, read top articles, and decide whether the evidence is sufficient.
- Add freshness and authority signals to ranking so current, canonical articles win over stale drafts.
- Return structured output with answer, citations and a confidence indicator, and escalate when confidence is low.
- Log every turn and review the unanswered and low-rated questions weekly.
- Feed gaps to content owners, publish fixes, and re-index so the assistant improves over time.
Try it yourself
Open the best model for agents selector →
From knowledge base to knowledge agent
A knowledge base stores answers; a knowledge assistant produces them. The difference is an agent loop that can search, read, check its own evidence and decide whether to answer or hand over. That loop turns a passive archive into a service with measurable quality.
On Plugsky the mechanics are standard: pass a tools array to /v1/chat/completions, execute the tool_calls the model returns against your knowledge services, append the results and iterate. Streaming keeps the experience responsive while retrieval and freshness checks run.
Freshness, authority and gap signals
Similarity is not truth. Two articles can be equally close to a query while only one is current and authoritative. Build ranking that combines vector similarity, keyword match, updated date, article status and source authority, then let the agent read the top candidates and reason over them.
- Recency: prefer articles updated within their review cycle and warn when citing older material.
- Authority: canonical, owner-approved articles outrank drafts, tickets and chat transcripts.
- Conflict handling: when sources disagree, the answer should present both with links.
- Escalation: ambiguous, sensitive or low-confidence questions should route to a person with context attached.
Measuring deflection and knowledge gaps
Two metrics tell you whether a knowledge assistant is working: how many questions it resolves without a human, and how quickly unanswered questions become published articles. Optimise only for deflection and you reward confident wrong answers; measure both and the system improves the underlying knowledge too.
Track citation accuracy, resolution rate, escalation accuracy and repeat-question rate. Review low-rated answers alongside the retrieved passages to find whether the fault was retrieval, ranking or the article itself. Route routine lookups to plugsky-micro or plugsky-lite and reserve stronger models for multi-source synthesis. Keep deployment aligned to policy — region-locked planes for cloud, VPC, on-prem or air-gapped where content cannot leave your network.
Honest comparison
| Concern | Plugsky knowledge agent | Help centre search | Managed answer bot |
|---|---|---|---|
| Answer basis | Retrieved, verified passages with citations | Ranked articles | Vendor knowledge model |
| Freshness | Updated-date and status signals in ranking | Manual curation | Varies |
| Escalation | Low-confidence routing to humans | Support queue | Vendor flow |
| Model choice | 30+ models, route by task | None | Vendor catalogue |
| Gap tracking | Unanswered queries logged as content backlog | Search analytics | Vendor reporting |
Frequently asked questions
What makes a knowledge assistant different from a chatbot?
A knowledge assistant is constrained to answer from approved sources with citations, and it escalates when evidence is weak. A general chatbot answers from model memory, which is unsuitable for operational knowledge.
How do you keep answers current?
Store updated date and review status on every article, use them in ranking, and re-index on publish. The agent should warn or escalate when only stale material exists.
Is function calling available?
Yes. OpenAI-style function calling and streaming are live on chat completions, which is all the retrieval loop needs.
How should the assistant handle sensitive questions?
Define escalation rules in your tools: anything touching legal, health, security or personnel matters routes to a human with the retrieved context attached.
Can we run it on-prem?
Yes. Plugsky supports region-locked cloud planes plus VPC, on-prem and air-gapped deployment, so knowledge content can stay inside your network.
How do we evaluate quality?
Score citation accuracy, resolution rate, escalation accuracy and repeat-question rate on a fixed question set, and review low-rated answers against their retrieved passages.
What is the fastest way to start?
Point the agent at one authoritative corpus, add search and escalation tools, and evaluate on real questions. The free plan's plugsky-micro and plugsky-lite models are enough to build the loop.