How Exa Vantage works, end to end.

01

Ingest

The expensive route is guarded before any model runs, then opens a streaming response.

• route handlerGET /api/report/stream validates the query (2-160 chars), enforces a per-IP rate limit (4/min, 40/day, fail-closed), and checks a global hourly/daily budget kill switch.
• streamIt opens an NDJSON ReadableStream (application/x-ndjson, no buffering) and hands a send() callback to the pipeline.

02

Route + junk-name filter

Gemini decides what you typed, screens out non-companies, and resolves the rest to a real company, with a cache shortcut on both ends.

• GeminirouteInput() first judges whether the input is a real research subject at all. A clearly non-company input (a personal name, a public figure, gibberish) is rejected here, before any cache or Exa work, emitting invalid so junk stops in ~1s instead of building a deck. The filter is high-precision: it accepts any plausible brand, even unrecognized or chef-named ones (e.g. Peter Chang), and any coherent sector.
• GeminiFor valid inputs the same pass classifies company vs sector and resolves the official name + domain; Exa search is the fallback when the model has no reliable domain. Emits meta.
• cacheA Firestore lookup by normalized input then returns a saved deck instantly (unless you hit Regenerate). A second check runs after the name resolves, to catch alternate phrasings.

03

Discover

Two Exa calls, run in parallel, build the candidate universe so we never over-index on a brand token.

• Exa contentsfetchSiteContent() reads the anchor's own site to ground everything in what it actually does.
• Exa ×2findSimilar off the seed domain and a semantic search run together; mergeHits() dedupes them. The market-size search is also kicked off here to overlap its latency.

04

Gate

Two Gemini gates keep only companies that are both genuinely relevant and actually acquirable.

• GeminifilterRelevant() drops name-collisions (the “Exa” vs “Exaforce” problem), judging by what each company does.
• GeminifilterIndependent() is the first independence layer: it drops sub-brands of a larger parent and the anchor's own subsidiaries, using model world knowledge. Both gates fall back to keeping the set if they would over-filter.

05

Synthesize

The surviving set is enriched into tear sheets, fact-checked against the live web, and sized, with a second independence pass on real evidence.

• Exa AgentAn emerging pass adds the off-database long tail: agentDiscoverEmerging() (Exa Agent deep research) when enabled, falling back to a freshness-tuned search. New names are deduped and capped.
• GeminisynthesizeLandscape() clusters everything into 3-6 segments and writes per-company tear sheets in one pass. Emits segments.
• ExafetchCompanyIntel() pulls facts + a recent signal per company (bounded concurrency of 6).
• GeminiextractQuant() is one batched pass that reads founding year, funding, stage, headcount, region, and a parent-owner signal, evidence-only. Any company with a detected parent is dropped here, the second independence layer, before it ever streams. Emits each company.

06

Position & persist

A market stat and a strategic thesis turn the set into a recommendation, then the deck is saved as a shareable link.

• Exa + GeminifetchMarketContext() searches the sector (not the company), ranks credible sources first, and extracts one stat whose source URL must match a retrieved page, or omits it. Emits market.
• GeminianalyzeOpportunity() writes the DealThesis: recommendation, conviction, why-now, tiered targets, value levers, risks, the ask, and a per-slide takeaway. Conviction is a derived, calibrated band (convictionSignal()) computed from the set, not a free model choice, and the argument is written to match it. Emits analysis then summary.
• FirestoresaveReport() persists the deck with cache keys; Regenerate overwrites the same /r/[id] so share links stay valid. Emits done.