How do you build sales agents with RAG?

Key facts

Pattern	Index approved content → retrieve evidence → cited drafts
Embeddings	plugsky-embed and plugsky-embed-multilingual via /v1/embeddings (live)
Generation	Draft replies with JSON mode and citations (live)
Metadata	Product, region, deal stage and content owner per chunk
Sources	Docs, battlecards, case studies and CRM notes
Models	30+ models behind one endpoint, route per task
Deployment	Region-locked planes, VPC, on-prem and air-gapped options
Free tier	plugsky-micro and plugsky-lite on the free plan, no card required

TL;DR

• Index only approved content — RAG amplifies whatever quality you feed it.
• Tag region and product metadata so answers match the buyer's context.
• Cite sources in every draft so reps can verify claims in seconds.
• Refresh the index when pricing, product or positioning changes.
• Keep humans in the loop; the agent prepares, the rep sends.

How it works, step by step

1. Assemble the approved corpus: product docs, battlecards, case studies, security answers and CRM notes with owner metadata.
2. Chunk by section and keep product, region, deal stage and owner in metadata for filtering.
3. Embed with plugsky-embed, or plugsky-embed-multilingual when content and prospects span languages.
4. Build retrieval that resolves the deal context and applies region and product filters before ranking.
5. Generate draft replies constrained to retrieved passages, with a sources array in JSON mode.
6. Route drafts to reps for edit and approval, and store the citations with the sent message.
7. Review unanswered questions monthly and feed them back to content owners.

Try it yourself

Open the RAG sandbox →

Grounding sales content in approved material

Sales conversations fail on accuracy: a wrong security answer or an outdated price costs credibility. RAG addresses that directly by grounding every draft in content the business has approved, with citations that let a rep verify before sending. The model's role shrinks to retrieval-aware writing, which is precisely where it is reliable.

Plugsky supplies the live model layer: /v1/embeddings for indexing and /v1/chat/completions for grounded drafts with JSON mode. Content, filters and permissions remain in your systems, which is where approved messaging governance belongs.

Indexing docs, case studies and CRM notes

Not everything deserves equal weight. Tag content with product, region, deal stage and owner so retrieval can prefer the most relevant and authoritative passage. Case studies suit late-stage conversations; security and compliance answers matter early in enterprise deals.

• Freshness: store updated dates and demote superseded pricing or positioning content.
• Regional variants: keep region metadata so local pricing and availability win.
• Customer notes: index CRM notes with strict access labels, or exclude them from shared retrieval.
• Keywords too: keep lexical search for SKUs, certification names and exact product terms.

Accuracy, approvals and evaluation

Drafts should carry their sources, and reps should be able to open each cited passage. When the corpus cannot answer a question, the agent should say so and route it to a product or security expert — a visible gap is more valuable than a confident guess in a live deal.

Measure answer accuracy against a question set with known correct responses, citation correctness, edit distance on drafts and the share of questions the corpus cannot answer. Those last two point directly at content work: missing articles and unclear positioning. Refresh the index whenever pricing or product changes, and deploy in a region that satisfies your customer data policy, including on-prem for restricted accounts.

Honest comparison

Capability	RAG sales assistant	Shared drive search	Generic chatbot
Answer basis	Approved content with citations	Documents found manually	Model memory
Region accuracy	Metadata filters per market	Manual	Unreliable
Freshness	Re-index on content change	Whatever is saved	Training cutoff
Draft quality	Grounded in retrieved passages	Rep writes	Plausible but unverified
Audit	Citations stored with sent messages	None	None

Frequently asked questions

Can the agent quote prices?

Only from approved content that you index, and regional pricing rules should be applied as filters. Never let a model recall prices from training data.

How do we handle security questionnaires?

Index approved security answers with product and certification metadata, and have the agent assemble a draft citing each answer so the security team can review quickly.

Should CRM notes be indexed?

Only with strict access labels and careful scoping. Many teams keep notes out of shared retrieval and use them in rep-scoped views instead.

Which embedding model fits global sales teams?

plugsky-embed-multilingual when content and prospects mix languages, so a question in one language can retrieve approved material in another.

How do we keep content current?

Re-index on publish, store updated dates, and demote superseded pricing or positioning content. Stale answers are the fastest way to lose a rep's trust.

How do we evaluate it?

Use a question set with known correct answers, score citation accuracy and edit distance on drafts, and track how many questions the corpus cannot answer.

Can it run on-prem?

Yes. Plugsky supports region-locked cloud planes plus VPC, on-prem and air-gapped deployment, so sensitive account data can stay inside your network.