Skip to main content
Status: Planned Tier: Pro

​
Overview

High IQ already knows a lot about you. It knows every strain you have purchased, which ones you rated highest, what terpene profiles dominate your collection, which High Families you gravitate toward, and how your preferences have evolved over time. The strain database contains 5,226+ profiles with embeddings that capture chemical similarity at a level far beyond “sativa vs. indica.” All the pieces are in place for deeply personalized recommendations. Smart Suggestions is the feature that connects these pieces. Instead of browsing the entire strain catalog and guessing, you open Smart Suggestions and see a curated list of strains selected specifically for you — based on your history, your terpene fingerprint, your effect preferences, and the gaps in your experience. “You loved Blue Dream’s myrcene-forward profile. Try OG Kush for a similar but deeper body experience.” “Your collection is heavy on Relaxing High strains. Here is an Uplifting High strain with a terpene profile you would probably enjoy based on your limonene affinity.” This is not generic recommendation. It is cannabis intelligence that gets smarter as you use it.

​
What It Does

  • Personalized strain recommendations — AI-selected strains based on your unique consumption profile, updated as your preferences evolve
  • Terpene fingerprint matching — Identifies your preferred terpene ratios from your collection and finds strains with similar profiles
  • High Family exploration — Suggests strains from High Families you have not explored, chosen for the highest likelihood of enjoyment based on your existing preferences
  • Effect-based discovery — “You frequently rate sessions as ‘focused’ and ‘creative’ — here are strains optimized for those effects”
  • Gap analysis — Identifies blind spots in your cannabis experience: “You have never tried a terpinolene-dominant strain. These are the most approachable ones for your profile”
  • Contextual suggestions — Time-of-day and day-of-week recommendations based on your session patterns: “For your typical Friday evening session, these strains perform best in your history”
  • Similarity explanations — Every recommendation includes a plain-language explanation of why it was chosen: “73% terpene overlap with your #1 rated strain, with additional caryophyllene that matches your preference for anti-inflammatory effects”
  • Confidence scores — Each suggestion includes a match confidence percentage so you know how strongly the algorithm believes you will enjoy it
  • Feedback loop — Rate suggestions as “interested” or “not for me” to continuously refine the algorithm
  • Wishlist integration — Add any suggestion to your wishlist with one tap

​
User Value

The “aha moment” is when Smart Suggestions recommends a strain you have never heard of, you try it, and it becomes your new favorite. That is the moment you trust the algorithm — and the moment High IQ becomes indispensable.

​
How It Works

1

Build Your Profile

Smart Suggestions activates after you have logged at least 5 strains in your stash or collection. The more data, the better the recommendations. A progress indicator shows: “Log 2 more strains for your first suggestions.”
2

Profile Analysis

The system analyzes your strain collection to build a terpene fingerprint (your preferred terpene ratios), a High Family distribution (which families you prefer), and an effects profile (which effects you seek).
3

Candidate Generation

The recommendation engine queries the strain database using your terpene fingerprint as an embedding search, filtered by High Family preferences and effect alignment. Hundreds of candidates are scored.
4

Ranking and Explanation

Candidates are ranked by a composite score combining terpene similarity, High Family alignment, effect match, novelty (strains you have not tried), and availability signals. The top 10-20 are selected, each with a generated explanation.
5

Presentation

The Smart Suggestions screen shows your recommendations as cards with strain name, High Family badge, match confidence percentage, key terpenes, and the explanation. Swipe right to add to wishlist, swipe left to dismiss.
6

Refinement

Every interaction (view, wishlist add, dismiss, purchase, session rating) feeds back into the model. Recommendations improve with every use.

​
Technical Approach

​
Recommendation Architecture

Smart Suggestions operates in three layers: Layer 1: Terpene Fingerprint (Collaborative Filtering) Your collection of strains defines a terpene fingerprint — a weighted vector of terpene preferences derived from the strains you own, weighted by your ratings and usage frequency. This fingerprint is compared against all 5,226+ strain embeddings using cosine similarity to find the chemically closest matches. Layer 2: High Spectrum Alignment (Content-Based) Your High Family distribution (e.g., 40% Relaxing, 30% Uplifting, 20% Relieving, 10% other) shapes the recommendation mix. The algorithm balances familiar territory (more strains from your preferred families) with gentle exploration (occasional suggestions from underexplored families with high predicted compatibility). Layer 3: Effect Optimization (Session Data) When session journal data is available, the algorithm incorporates your subjective outcomes. Strains that produce effects you rate highly are weighted more heavily. Strains associated with negative effects (anxiety, headache) in similar terpene profiles are penalized. This layer only activates after sufficient session data (20+ sessions).

​
Scoring Formula

score = (terpene_similarity * 0.4)
      + (family_alignment * 0.25)
      + (effect_match * 0.2)
      + (novelty_bonus * 0.1)
      + (availability_signal * 0.05)
Weights are tuned based on user feedback signals and will evolve as we collect data on recommendation accuracy.

​
Explanation Generation

Each recommendation includes a natural-language explanation generated by Professor High’s voice. Examples:
  • “This strain shares 78% terpene overlap with Gorilla Glue, your top-rated strain, but adds a pinene kick that you might enjoy based on your collection trends.”
  • “You have been gravitating toward Relaxing High strains lately. This Entourage High strain bridges your comfort zone with a more complex terpene profile.”
  • “Your session data shows you rate caryophyllene-heavy strains 0.8 points higher on average. Wedding Cake is loaded with it.”

​
Infrastructure

ComponentTechnologyStatus
Strain embeddingsExisting vector embeddings in SupabaseBuilt
Cosine similarity searchSupabase pgvectorBuilt
User collection dataConvex (stash, favorites, ratings)Built
Terpene fingerprint calculationNew Convex queryPlanned
Recommendation API endpointHono APIPlanned
Explanation generationAI SDK + Professor High promptPlanned
Feedback storageConvex mutationsPlanned

​
Tier Impact

TierAccess
FreeNot available
ProFull personalized recommendations, explanation cards, wishlist integration, feedback refinement, contextual suggestions

​
Dependencies

  • Strain embeddings and similarity matching — built and live
  • High Spectrum classification for all strains — built and live
  • User stash and collection data in Convex — built and live
  • Strain ratings and favorites — built and live
  • Terpene fingerprint calculation from user collection
  • Recommendation scoring engine (API endpoint)
  • Explanation generation with Professor High voice
  • Smart Suggestions UI (card list with swipe actions)
  • Feedback loop (interested/dismiss signals stored and weighted)
  • Session journal integration for effect-based optimization (planned)

​
Open Questions

  1. Cold start problem — With only 5 strains, recommendations will be rough. What is the minimum collection size for useful suggestions? Should we use onboarding preference data (favorite effects, preferred High Families) as a seed until the collection is large enough?
  2. Refresh frequency — How often should the suggestion list refresh? Daily feels right (new strains to consider each morning), but recalculating for every user daily may be expensive. Weekly refresh with manual “refresh” button?
  3. Availability awareness — Should recommendations factor in what is actually available at dispensaries near the user? This massively increases value but requires dispensary menu data integration. Defer until Dispensary Map is built?
  4. Explanation depth — How much chemistry should the explanations include? “78% terpene overlap” is precise but nerdy. “Very similar to your favorite” is accessible but vague. Let users set their preferred explanation depth (casual, balanced, scientific)?
  5. Negative recommendations — Should we explicitly warn about strains that might cause effects the user dislikes? “Based on your history, avoid limonene-dominant strains in the evening” is valuable but could feel paternalistic.
  • Session Journal — Provides the subjective outcome data that supercharges recommendations
  • Health Sync — Adds objective health outcome data to the recommendation model
  • Strain Discovery — The existing search and browse experience that Smart Suggestions enhances
  • Shopping Agent — Combines recommendations with availability: “This is your best match AND it is in stock 2 miles away”
  • Collections & Rankings — User ratings and collections are the primary input to the recommendation engine