[FEAT] LLM Rag pipeline for Recommendation

[Jrodrigo06]

---

name: LLM Rag pipeline for Recommendation
title: "[FEAT] - Rag Recommendation Pipeline"
labels: feature, backlog
assignees: @mdeekshita @SHarg9876 @keshavgoel787

Summary

LLM-powered RAG pipeline that auto-suggests ingredients and trigger bucket tags

Motivation

Likely that users are unaware of buckets and potential ingredient triggers, so suggestions can be a really powerful feature!

Requirements

Acceptance Criteria

Sub-task 1: Tag & Bucket Models

• Tag model created with fields: id (UUID), name (String), is_system (bool — True if LLM-suggested, False if user-defined), created_at
• FoodLogTag join table created linking food_logs.id → tags.id
• System tags seeded for core trigger buckets: gluten, FODMAPs, nightshades, histamines, added sugar, artificial additives, dairy, FDA Big 9 allergens
• TagCreate and TagResponse schemas created
• SuggestedIngredientResponse schema created with fields: name (String), buckets (list of String)
• FoodLogResponse updated to include suggested_tags and suggested_ingredients: list[SuggestedIngredientResponse] fields

Sub-task 2: Tag Endpoints

• POST /food-log/{food_log_id}/tags — confirms and persists tags to a food log
• POST /tags — creates a custom user-defined tag (is_system=False)
• GET /tags — returns all available tags for a user to pick from
• Unconfirmed tags must not be persisted until user explicitly confirms

Sub-task 3: Vector Store & RAG Infrastructure

• pgvector extension enabled on Supabase
• KnowledgeChunk model created with fields: id (UUID), content (Text), embedding (Vector), source (String)
• Embedding generation utility using Gemini models/text-embedding-004
• Similarity search function that takes a food name query, returns top N relevant chunks from the knowledge store
• Seeding script scaffolded with TODOs for document sources (Monash FODMAP, FDA Big 9, AIP framework) — actual content to be added once sources are confirmed

Sub-task 4: RAG Tagging Service

• RAGTaggingService created that follows the same constructor pattern as FoodService
• Given a food name and its ingredients JSONB field, runs similarity search to retrieve relevant trigger context chunks
• LLM prompt built with: food name, ingredients, retrieved context chunks — instructs LLM to suggest likely ingredients and classify each into trigger buckets without hallucinating unlikely ingredients
• Returns structured response: suggested_ingredients: list[SuggestedIngredientResponse]
• Hooked into food log creation flow — suggestions returned in response, not auto-saved

Out of Scope

• Frontend confirmation UI — backend only
• Actual seeding of knowledge documents — scaffolded only, content TBD
• Apriori analysis — separate ticket
• Barcode scan or external food API integration

Technical Approach

Affected Areas

• backend/models/tag.py (new) — Tag and FoodLogTag models
• backend/models/knowledge_chunk.py (new) — vector store model
• backend/models/food_log.py — add relationship to FoodLogTag
• backend/schemas/tag.py (new) — TagCreate, TagResponse, SuggestedIngredientResponse
• backend/schemas/food.py — add suggested_tags and suggested_ingredients to FoodLogResponse
• backend/services/tagging_service.py (new) — tag business logic
• backend/services/rag_tagging_service.py (new) — RAG + LLM suggestion logic
• backend/repositories/food_repository.py — add tag persistence methods
• backend/repositories/tag_repository.py (new)
• backend/routers/food_router.py — update create endpoint
• backend/routers/tag_router.py (new) — register in main.py

Dependencies

• Food, FoodLog models exist
• FoodService, FoodRepository patterns exist
• pgvector must be enabled on Supabase before sub-task 3 begins (I might have to update the docker image for the test db for this I'll look into this though)

Testing Notes

Follow existing integration test patterns in backend/tests/integration/

• LLM suggestion — given a Food with populated ingredients JSONB, correct trigger bucket tags and suggested ingredients returned
• Sparse ingredients — Food with ingredients=None should return reasonable suggestions from food name alone without hallucinating
• RAG retrieval — similarity search should return relevant chunks for a given food query
• Tag confirmation — unconfirmed tags must not appear in DB after food log creation
• Suggested ingredients — must never be persisted, only returned in response
• Custom tag creation — is_system=False tag persists and appears in GET /tags
• Duplicate tags — saving same tag to a food log twice should not create duplicate join table entries

Additional Context

(https://www.youtube.com/watch?v=T-D1OfcDW1M)