Use Cases + Implementation

How do you build support agents with RAG?

A RAG support agent retrieves from your help centre, release notes and resolved tickets before it answers, then cites what it used. Embed the knowledge base, filter retrieval by product, version and customer entitlements, and pass the best passages to a chat model that is instructed to answer only from context. Deflect repeat questions, and escalate with the retrieved evidence attached when confidence is low.

Key facts

EmbeddingsLive embeddings API for help-centre chunks with metadata filters
GenerationOpenAI-compatible chat completions with streaming and JSON mode (live)
PipelineIngest, chunk, embed, permission-filter, retrieve, answer with citations
Models30+ models; cheap tiers for FAQ retrieval answers, stronger tiers for complex cases
FreshnessStore document versions with vectors so stale articles can be expired
PricingFlat monthly self-serve plans with no per-token charges on self-serve
Free tierplugsky-micro and plugsky-lite on the free plan, no card required
RoadmapFiles and batch endpoints are coming soon; ingest documents in your own pipeline today

TL;DR

  • Retrieve from help centre, release notes and past tickets before answering.
  • Filter by product version and entitlements, not just tenant.
  • Cite the article or ticket behind every answer so agents can verify it.
  • Deflect repeat questions; escalate with evidence instead of guessing.
  • Re-index on content changes and measure answer accuracy monthly.

How it works, step by step

  1. Collect the knowledge sources: help articles, macros, release notes and resolved tickets with their resolutions.
  2. Chunk articles on headings and ticket threads on turns, and attach product, version and visibility metadata.
  3. Embed chunks with plugsky-embed and store vectors plus source links in your database.
  4. At question time, embed the customer message, filter by entitlement and product, and retrieve top candidates.
  5. Pass the passages to a chat model with an answer-only-from-context rule and a citation format.
  6. Add a confidence check: when retrieved evidence is weak, route to a human with the citations attached.
  7. Track deflection rate, citation accuracy and reopened tickets, then tune chunking and top-k.
1Collect theknowledge sources:help articles,2Chunk articles onheadings and ticketthreads on turns,3Embed chunks withplugsky-embed andstore vectors plus4At question time,embed the customermessage, filter by5Pass the passagesto a chat modelwith an6Add a confidencecheck: whenretrieved evidence

Try it yourself

Open the best-model-for-RAG selector →

Retrieve first, answer second

Support answers live in your documentation, and documentation drifts. The agent's job is to find the current, correct passage and present it, not to recall policy from training data. Ingest help articles, macros, release notes and resolved tickets; chunk them so each unit is a coherent instruction or resolution; embed with plugsky-embed; and store the source link with every vector so answers can cite their origin.

Retrieval quality is decided before the model speaks. Keep product, version and visibility in metadata so a customer on an old plan never sees instructions for the new one. Deduplicate near-identical articles during ingestion, because redundancy crowds the prompt with the same advice three times.

Ticket context and citations

The customer's own history is part of the context: plan, products owned, recent tickets and the conversation so far. Assemble it compactly, then let retrieval supply the procedural knowledge. Ask for citations in a structured shape — JSON mode works well — so the interface can render links next to each claim.

  • Resolve citations to real URLs or ticket ids before showing them; broken references erode trust.
  • Store the document version with each chunk so answers can be traced to what was live at the time.
  • Keep the last few turns verbatim and summarize older history instead of resending it all.

Deflection, escalation and freshness

A useful support agent does two things well: it resolves the repeatable questions instantly, and it exits gracefully when the question needs a person. Track deflection by intent rather than as a single number, because a high aggregate can hide poor performance on billing or security questions where errors are expensive.

Schedule re-indexing when content changes and expire vectors for retired articles, otherwise the agent will confidently quote last year's policy. Evaluate monthly on a set of real questions with known answers, checking citation accuracy and refusal quality alongside resolution. The free plan with plugsky-micro and plugsky-lite is enough to build the pipeline; the 14-day full-access trial lets you test stronger tiers on your hardest tickets, and the live pricing page covers paid options.

Honest comparison

ConcernRAG support agent on PlugskyPrompt-only chatbotHuman-only support
Answer sourceRetrieved help articles and ticketsModel memoryAgent knowledge
CitationsStructured links to sourcesNoneClickable docs in replies
FreshnessRe-index on content changeStale until retrainedTraining dependent
Entitlement safetyMetadata filters before rankingNot enforceableManual checks
Cost profileCheap tiers for common questionsUnbounded confident answersHeadcount

Frequently asked questions

What sources should a support RAG agent use?

Your help centre, macros, release notes and resolved tickets with their resolutions. Chunk each source so a unit expresses one instruction or fix, and keep the source link in metadata for citations.

How do I keep answers current?

Re-index when content changes and store document versions with vectors so retired articles can be expired. Stale vectors are the most common cause of confidently wrong support answers.

Should retrieval respect customer entitlements?

Yes. Filter by product, version, plan and visibility before ranking, so a customer never receives instructions for a feature they cannot use.

Which embedding model should I use?

plugsky-embed for English-heavy help centres and plugsky-embed-multilingual when content or customers mix languages. Embed documents and questions with the same model.

How does the agent know when to escalate?

When retrieval returns weak evidence, when the question matches a sensitive category, or when the customer asks for a person. Escalate with the retrieved citations attached.

Can answers include links?

Yes. Request a structured output mapping claims to chunk ids, resolve those ids to real URLs in your application, and render them beside the answer.

How do I measure success?

Track deflection per intent, citation accuracy, reopened tickets and refusal quality on a set of real questions, and re-run after every content or prompt change.

Can I try it for free?

Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and a 14-day full-access trial is available for stronger models.