@absolutejs/partnership

In-house partnership & relationship intelligence for AbsoluteJS apps — AI reasoning over the partnership lifecycle, so you don't pay a deal-copilot SaaS per seat for the parts you can run yourself.

Each primitive is a pure function of typed input + an injected AI call. The package owns the prompt, the JSON Schema, the tool contract, and the result mapping. Your app supplies the generation call — and with it the provider, billing/metering, timeouts, and any fallback. The package never imports a provider and never touches your ledger.

Install

bun add @absolutejs/partnership

Wiring (one-time)

Bridge the package's generateObject to whatever you already use for structured generation. If you use @absolutejs/ai (directly or behind a metering wrapper), it's one line — the request shape is the subset of generateObjectAI's options the package controls, minus provider and your metering fields:

import type { GenerateObject } from "@absolutejs/partnership";
import { meteredGenerateObjectAI } from "./usage/meteredAI";
import { aiProvider } from "./integrations/aiProvider";

export const partnershipCtx = (userSub?: string | null) => ({
  generateObject: ((req) =>
    meteredGenerateObjectAI({ ...req, provider: aiProvider, userSub })) as GenerateObject,
});

The package passes its own feature tag through ("trustFit", "classifyMatch", "personConnection", "verification"), so your usage ledger keeps its existing per-feature breakdown.

Primitives (Wave 1A)

import { scoreTrustFit } from "@absolutejs/partnership";

const { dimensions, reasons } = await scoreTrustFit(
  {
    member: { niche: "devtools", offer: "hosted CI", audienceSize: "10k" },
    partner: { company: companyData, person: personData, reasoning },
    priorScores: { theirReceptiveness: 0.4, yourFit: 0.9 },
  },
  partnershipCtx(userSub),
);

Web-grounded primitives

verifyPartner optionally takes an injected research call (e.g. your @absolutejs/discover web-research) on the context. Omit it and synthesis falls back to knowledge-only:

await verifyPartner(input, {
  ...partnershipCtx(userSub),
  research: (system, user) => webResearch(system, user),
});

Fallbacks

Primitives throw on a model or validation failure. Heuristic fallbacks (e.g. deriving rough Trust & Fit dimensions from a match's existing signals) are app-specific, so they stay in your wrapper — catch and degrade however your UI needs.

License

BSL 1.1 → Apache 2.0 on the Change Date (see LICENSE). You may build and ship your own apps and SaaS on top of it; you may not offer it as a competing hosted partnership-intelligence / deal-copilot service.

Research evidence (0.1)

ResearchFn accepts structured research alongside legacy string | null. Verification, competitor analysis and meeting preparation retain researchStatus, sources, findings and limitations. researchedAt is empty when source-bound research was unavailable; generatedAt is always generation time. Existing string research remains explicitly unverified.

Structured research invokes a separate researchEvidenceReview generation through the host's generateObject, which must be budgeted and metered like the initial synthesis. fieldEvidence binds reviewed output paths to exact source excerpts, supplied input, proposed assessments or unresolved facts. Unsupported factual text becomes unresolved, and corporate affiliations are removed unless every textual field is source-supported. analysisStatus describes that review; numeric scores remain interpretations, not verified facts. A model's entailment judgment still requires human review for consequential decisions.

Partner Revenue Pipeline

calculatePartnerRevenuePipeline (also in the browser-safe @absolutejs/partnership/revenue entry) computes company revenue divided by annual recurring pricing as customer equivalents, then multiplies by the client's annual pricing as a theoretical annual revenue opportunity. This assumes 100% conversion and is not a customer count or revenue forecast. Results retain evidence, assumptions, limitations, and low scenario confidence.

Supply partnerAnnualRevenue, partnerPricing, and clientPricing as RevenueAmount objects (or null). Each amount includes amount, ISO currency, period (month or year), billing unit (company, customer, or seat), and evidence (basis, sourceTitle, exact quote, ISO observedAt, confidence, and source URL for documented claims). Pricing units and currencies must align. Seat prices cannot be used as customer prices. Usage-based businesses require a sourced or client-reported yearly customer value with pricingModel: "annual_customer_value"; do not multiply a per-call price by twelve.

researchPartnerRevenuePipeline(input, researchContext) uses the host's injected, metered research and structured-generation calls. Wire Brave-backed research through the existing ResearchContext.research contract. Only structured, retained source quotes with matching numeric amounts can populate research inputs. Unknown revenue, incomparable business scope, missing prices and unsupported evidence remain unavailable. Research cannot guarantee a public revenue disclosure exists. No foreign-exchange conversion is implicit.

When clientPricing is missing, supply clientName and optionally clientWebsite to research the client's public recurring pricing alongside partner revenue and pricing (three query intents). Explicit client pricing always wins and skips the extra query. Usage-based prices remain unknown unless there is an explicit source-supported annual customer value.

Source-backed asset construction (0.3)

draftPartnershipAsset and editPartnershipAsset return evidenceDocument with the title and Markdown. Persist it with the immutable input sources. Source blocks select whole records by content-addressed ID. The server renders their exact text and provenance. Announcement statement blocks instead contain model-authored prose linked to those original records; citation integrity is not proof that the prose is true. validateAssetEvidenceDocument(sources, markdown, evidenceDocument) rejects changed sources, citations, plans or output. Call it before publication/export. Existing prose is editing context, never new evidence.

Source quotations are reported information, not verified truth. Intake summaries can themselves contain AI errors. Hosts must distinguish primary records from summaries and retain their source snapshots. Sources are escaped as inert text; their content cannot choose its own provenance classification.

Proposal and question blocks remain model-authored, visibly labeled, and must pass a host-side review rejecting any factual assertion or presupposition. Construction alone cannot establish semantic truth or appropriate proposals. calculateAssetScenario computes a labeled scenario from complete numeric source fields, matching annual periods/currencies, and explicit assumptions; it rejects invented amounts and never represents potential revenue as guaranteed.

Evidence eligibility

A stored AI analysis is not primary evidence. Match sources are context-only by default. Hosts may set evidenceSelection to exact JSON-pointer fields and optional decoded-text start/end spans. Use it to select original user turns, supplied documents, or independently retained evidence; never grant eligibility to generated reasoning simply because it is stored. An empty selection produces no factual records. Selection metadata is included in the construction digest, and validation rejects old documents that used excluded records. Offsets still address the original immutable decoded source value. This establishes provenance, not the truth of a source or completeness of a model's selection.

Reader-facing source-backed documents

Use presentAssetEvidenceDocument(sources, storedMarkdown, evidenceDocument) for exported documents and their previews. It validates the stored construction first, then produces clean Markdown with numbered sources, reader-facing source labels/excerpts, and separate missingInputs. Storage paths, hashes and internal reference IDs remain in the immutable audit document instead of appearing in the asset. Keep the original Markdown/proof for validation; do not replace it with presentation Markdown. Preview the same rendered file that downloads use rather than building a second document UI.

Document readiness and authored announcements

Run reviewAssetContent({ assetType, sources, document, markdown }, ctx) before marking a draft ready. Every selected block is reviewed in its section and format context. assetContentIssues(review) includes misplaced or stale content, missing required facts, and document-level failures. Persist the review bound to the source/document digests and asset type; reuse it only while those inputs and the review policy remain unchanged. Missing essential facts should request member input instead of triggering a paid rewrite that cannot supply them. Optional quotes, secondary contacts and marketing detail should not block readiness.

For statement blocks, independently check every assertion against the referenced original source excerpts. onSpark performs a separate entailment pass and the whole-document review before publication. Reject changed uncertainty, targets presented as achieved outcomes, invented agreement, and any unsupported detail; valid source IDs alone must never pass that gate. The shared document review includes statement excerpts for the same entailment check. It sends selected evidence only, not unrelated intake history. This is semantic review, not a guarantee against all model errors or independently verified source truth.

Legal generation permits only supplied source terms or questions, not invented proposal defaults. Generation also classifies the intended use of sources to reject objectives/baselines under Results, stale correction text, and business background used as operative legal terms. The temporary source-use classification is not added to the immutable evidence format.