Grounded, not generated
Every claim the operator makes traces to a Perplexity citation. If a human asks "where did that come from," the answer is one MongoDB query.
Stack guide
The Perplexity research-agent stack for commercial operators
Perplexity is excellent at finding things on the live web. The job of the surrounding stack is to turn that into evidence operators can act on — with citations, recency checks, and an audit trail.
The stack
Updated · 2026-05-21
01TL;DR
02The stack
- L/01Retrieval surface
Perplexity does the live-web search and returns answers grounded in citations the operator can verify.
- L/02Reasoning core
Anthropic Claude takes Perplexity's output and decides what to do with it — file a brief, update a record, escalate a flag.
- L/03Tool gateway
Glama exposes Perplexity as one MCP tool. Caching and rate-limiting happen at the gateway, not in the prompt.
- L/04Memory + audit
Supabase caches recent research per account to keep token cost down. MongoDB stores the citation set behind every action the operator took.
- L/05Human review surface
Critical research outputs land in an Obsidian vault humans can search, link, and annotate. The vault is the operator's working memory humans can audit at their own pace.
03Why this stack
Recency control
The operator queries fresh every time something matters; cached only when it doesn't. Glama enforces the policy so the model doesn't have to remember.
Cost discipline
Supabase caching cuts redundant Perplexity calls during a single workflow run. Token cost stays linear with new information, not with operator chatter.
Reviewable trail
Obsidian vault means a human can read the operator's research notes the way they'd read another analyst's. Trust grows because the work is legible.
04Where it shines
- ◇/01
Competitive intelligence operators that brief sales teams nightly
- ◇/02
Compliance and regulatory-update operators that watch a jurisdiction list
- ◇/03
Prospect-research operators that enrich CRM records before outreach
- ◇/04
Procurement and price-monitor operators that flag market moves
05Comparison
Perplexity-centric research stack
Pros
- · Live web grounding with citations
- · Cached and rate-limited via Glama
- · Human-readable vault in Obsidian
Cons
- · Sources are public web — paywalled databases need a separate adapter
Direct web search (Bing / SerpAPI) with custom retrieval
Pros
- · Maximum control over ranking and source filtering
Cons
- · You build the synthesis layer yourself — months of work
- · Citation hygiene becomes your problem
Pure RAG over internal docs
Pros
- · No external dependencies
Cons
- · Can't answer anything outside what you've already indexed
- · Recency depends on your re-indexing cadence
06Implementation notes
- 01
Set per-workflow freshness policies. Competitive price data needs sub-hour recency; regulatory monitoring may tolerate a daily refresh.
- 02
Use Glama's rate-limiting to keep Perplexity spend predictable. A runaway operator querying every minute is the most common cost surprise.
- 03
Cache Perplexity results in Supabase keyed by the canonicalised query plus a TTL. The operator only re-queries when the TTL expires or when the workflow demands fresh.
- 04
Pipe every cited URL into the MongoDB log against the action it informed. "Why did the operator say X" should be one cross-join.
- 05
For Obsidian, structure the vault as one note per account, one section per research run. Humans navigate by account; auditors navigate by action.
- 06
Add a recency check in the reasoning prompt — "if the most recent citation is older than X days for this workflow, escalate."
07Related
Industries it fits
Workflows it fits
08Questions
Why pair Perplexity with Claude rather than letting Perplexity answer directly?
Perplexity is excellent at retrieval and decent at summarisation. It's not designed to take action — file the brief, update the record, queue the alert. Claude does the deciding; Perplexity does the finding.
How does the stack handle paywalled or proprietary sources?
Perplexity covers public web. For paywalled sources we add a separate adapter inside Glama (vendor APIs, licensed feeds). The operator sees them as additional tools with the same calling pattern.
What happens when Perplexity is wrong?
The reasoning prompt requires citations for every claim that drives action. If the citation doesn't support the claim, the operator escalates rather than acts. The MongoDB log shows both the citation and the operator's reasoning.
Can a human review the research before the operator acts?
Yes — that's the default for high-stakes workflows. The operator drafts the action, attaches the citations, and waits for human sign-off. For low-stakes routine workflows the action proceeds and the human reviews after.
How long does it take AIMOCS to deploy a research operator?
Three to four weeks. Week one is workflow + source mapping, week two is Glama wiring and cache policy, week three is operator-on-shadow against real queries, week four is staged handover with human review.
09Begin
We don't advise on AI. We run it for you.