Policy-Driven Compatibility Beyond 100 Parameters
Unlimited descriptive policies per user. RAG-powered semantic matching. Privacy-first architecture.
Each user creates policies at three tiers. Each policy is a plain-English document — as long and detailed as needed. ContextWeaver indexes them as RAG vectors and matches semantically.
Hard limits. If violated → instant reject.
Heavily weighted. Missing these lowers the match score significantly.
Bonus points. Additive — presence improves the score.
The power: Each policy can be a single sentence or a detailed paragraph. "I want someone who grew up with modest means and understands the value of hard work, but has the ambition to build something bigger — not someone born wealthy who's never had to struggle." Try expressing that with a checkbox.
index_policies(user="sarah", tier="dealbreaker", docs=12)index_policies(user="alex", tier="dealbreaker", docs=9)
Each policy document is chunked, embedded as a 3072-dim vector, and stored in a per-user private index. Sarah's policies are never visible to Alex and vice versa.
cross_match(sarah.dealbreakers → alex.profile)cross_match(alex.dealbreakers → sarah.profile)
score_preferences(sarah.strong_prefs → alex.profile, alex.strong_prefs → sarah.profile)
| Sarah's Preference | Matches In Alex? | Score |
|---|---|---|
| "Emotionally mature, can discuss feelings" | Alex values deep communication, mentions therapy positively | 94% |
| "Has genuine friendships" | Alex describes weekly game nights with college friends | 91% |
| "Ambitious but present" | Alex: "I work smart hours, evenings are mine" | 88% |
| Alex's Preference | Matches In Sarah? | Score |
|---|---|---|
| "Intellectually curious" | Sarah reads 2 books/month, takes online courses for fun | 96% |
| "Financially grounded" | Sarah: "I max out my 401k and have 6 months emergency fund" | 93% |
| "Values alone time" | Sarah: "I need my solo Sunday mornings — non-negotiable" | 97% |
| Feature | Traditional Apps | ContextWeaver |
|---|---|---|
| Parameters per user | 30–100 fixed fields | Unlimited — each policy is a rich document |
| Expressiveness | Checkboxes, dropdowns, ranges | Plain English paragraphs — any nuance |
| Matching method | Boolean + heuristic scoring | 3072-dim semantic vector similarity (RAG) |
| Priority weighting | Basic (if any) | 3-tier hierarchy: dealbreaker > preference > bonus |
| Bidirectional | Usually one-directional | Both sides cross-checked symmetrically |
| Privacy | Profiles visible to matches | Policies never exposed — only the match score |
| Evolving preferences | Manual profile edits | Add/remove policies anytime, re-index instantly |
| Explanation | "You matched!" — no reason given | AI explains why you matched on each dimension |
The same policy-based matching engine works for any two-sided marketplace:
Candidate policies: "Remote only, no surveillance tools, modern tech stack"
Company policies: "Must have K8s experience, growth mindset"
"Clean kitchen daily, quiet after 10pm, OK with a cat, splitting utilities evenly, no overnight guests on weekdays"
"Technical co-founder, bootstrapping mindset, willing to do 18 months without salary, B2B SaaS experience"
| Component | Role in Matching |
|---|---|
| Private Indexes | Each user's policies stored in their own vector index — never visible to others |
| Hierarchical RAG | Org policies (platform rules) → Group policies (community standards) → User policies (personal preferences) |
| Vault | Identity verification tokens, premium subscription status stored securely |
| Workflow Engine | Orchestrates the multi-step match: index → dealbreaker check → score → bonus → notify |
| MCP Tools | index_policies, cross_match, score_preferences, send_match_notification |
| Visual Designer | Platform operator designs the matching plugin — drag & drop connectors, policies, and notification flows |
| Multi-User Security | User A can never query User B's policy index. Only the match engine (org-level service) can cross-reference. |