Optional parallel track

Tool recommendation systems

Build a recommender that combines semantic similarity, graph structure, lexical evidence, intent, and telemetry, then test the design in production Snowflake.

2 artifacts · Level 2 → Level 3 · strict paired reading
Read the pair in order

This track sits alongside the 108-week spine rather than adding another phase. Read the theory first; the production artifact assumes its equations and makes their operational seams visible.

1 · Level 2 · theory

~ KB · numbered sections · ~2 hr full read

2 · Level 3 · production

~ KB · numbered sections · ~3 hr full read

The production seam

Where recommendation math becomes a platform

The same score needs graph-cycle guards, heterogeneous-score normalization, dynamic thresholds, incremental refresh, and a serving path that does not consume the warehouse budget. That is why the production reading is mandatory after the theory, not an interchangeable implementation note.

AI_EMBED + VECTORrecursive CTE + cycle guardweighted Adamic–AdarPERCENTILE_CONTStreams + TasksCortex Search
Recommendation pipeline from query signals through graph and embedding scores to explainable ranked tools
Cross-reference map

Each primitive below is already load-bearing elsewhere. The recommender track makes the combination explicit.

Spine / capstoneLoad-bearing primitiveWhy it connects
W31 cosine similarity + length-weighted pooling retrieval math
W36 IDF-weighted token scoring chunk and lexical scoring
W39 Recall@K, MRR, NDCG evaluation ranking metrics
W77 graph traversal + hop decay retrieval patterns
W78 dynamic noise-floor filtering retrieval evaluation
W81 Adamic-Adar for tool graph agent foundations
W86 operational telemetry into ranking monitoring
W94 multi-signal router traffic routing
W97 type affinity for tenant policies quotas and priority
Capstone 3 OpenRouter gateway as multi-signal recommender capstone integration
Optional depth, strict order

SEARCH.md can stand alone as general recommendation math. SNOWFLAKE.md is deliberately marked as a prerequisite-of-SEARCH pair member: do not start with its SQL. Log one checkbox per section in the learner scaffold and record the production failure you would monitor for each signal.

Return to the 108-week spine →