OUR MISSION

AI commerce for
regulated retail

ShopSupport connects to your POS and e-commerce APIs, trains agentic systems on your catalog and compliance policies, and deploys shopping assistants for dispensaries and wineries.

Integration-first, production-always

Dispensaries and wineries run on specialized POS systems — Dutchie, Treez, Shopify, custom APIs — with catalogs that change hourly and compliance rules that vary by state. Generic chatbots can't handle that complexity.

ShopSupport starts with your live data. We connect to your POS, ingest your catalog, embed your policies, and deploy agents that know what's in stock right now — not what was in stock when someone last updated a spreadsheet.

Everything we build gets evaluated against one question: does it help a shopper find the right product and complete a purchase? If the answer is yes, it ships.

Recent Research Signals Tracking
Self-RAG: Learning to Retrieve, Generate, and Critique
Retrieval · Adaptive reasoning
HyDE: Hypothetical Document Embeddings for retrieval
Embedding · Query expansion
RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval
Hierarchical chunking · Long docs
ReAct: Synergizing Reasoning and Acting in LLMs
Agent loops · Tool use

Towards intelligent commerce at scale

Every dispensary and winery shopper deserves expert guidance — strain recommendations, wine pairings, compliance-aware suggestions — without waiting for staff. That's what we're building.

Dispensary Commerce

Compliance-aware product discovery for cannabis retail. Strain recommendations with potency filters, purchase limit tracking, and live Dutchie or Treez inventory sync. Every suggestion is in stock and within regulations.

Winery DTC

Expert guidance from vineyard to checkout. Tasting notes, food pairings, club membership tiers, and shipping zone validation — all connected to your live catalog and DTC fulfillment rules.

POS-Native Intelligence

General-purpose LLMs don't know your inventory. ShopSupport agents query your POS in real time, enforce compliance policies, and reason over your specific catalog — strains, vintages, SKUs — like your best budtender or sommelier.

From your data to a working AI, practically

RAG (Retrieval-Augmented Generation) sounds complex. The practical version is straightforward: the AI retrieves relevant pieces of your knowledge before generating an answer. Here's what that looks like in production.

PATH A

You have an API

If your product data, knowledge base, or content is already served via a REST or GraphQL API, we connect directly. The AI queries your live data at inference time, meaning it always reflects the current state of your catalogue, pricing, or policies without any re-indexing.

Your API
Agent retrieves
Answer generated
Real-time · Always current · No indexing lag
or
PATH B

You have documents

PDFs, Word docs, support transcripts, policy manuals, product spec sheets. We ingest, chunk, embed, and index them into a vector store. Chunking strategy matters enormously: we use semantic boundaries, sliding windows, and metadata tagging so the retriever finds the right passage, not just the right document.

Documents
Chunk + embed
Vector search
Answer
Offline indexing · Semantic search · No API needed

Why chunking strategy matters

Chunking is where most RAG implementations fail silently. Split too aggressively and you lose context. The retrieved passage makes sense in isolation but misses the surrounding reasoning. Split too broadly and retrieval becomes imprecise. You get the right page but not the right paragraph.

We use a layered approach: semantic chunking at sentence boundaries, overlapping windows for context continuity, and per-chunk metadata (section title, document type, date) that can be used as a retrieval filter. A compliance query filters to policy documents. A product query filters to a specific category and potency range. The retriever finds not just semantically similar content, but the right kind of content.

Source document
§ 2.1
↗ retrieved
§ 2.2
§ 2.3
↗ retrieved
§ 3.1
hybrid search
Context window
Chunk § 2.1 (94%)
Chunk § 2.3 (81%)
Training Pipeline
01
Domain corpus collection
Gather high-quality in-domain text, clean and deduplicate
02
Instruction dataset construction
Format examples as prompt–completion pairs with domain-specific Q&A
03
Supervised fine-tuning (SFT)
LoRA or full fine-tune on base model; evaluate on held-out domain set
04
RLHF / DPO alignment
Preference tuning to align tone, refusal behavior, and domain accuracy
05
Eval + red-teaming
Benchmark against domain-specific metrics; probe for edge cases

Training AI into subject matter expertise

Retrieval can answer questions about your data. Fine-tuning shapes how the model reasons, responds, and represents your domain. The two approaches are complementary. Most production systems need both.

We work with LoRA (Low-Rank Adaptation) for efficient fine-tuning of large models without full parameter updates, making it practical to specialize a strong base model on domain-specific instruction data. For smaller models that need to run on-premise or with tight latency requirements, we handle full fine-tuning and quantization.

The result is a model that writes in your brand voice, knows your terminology, handles your edge cases, and refuses gracefully when a question falls outside its competence, rather than hallucinating a confident-sounding wrong answer.

SHOPSUPPORT.AI

Built for dispensaries and wineries.

We work with a focused set of regulated retail stores — connecting to your POS, training on your catalog, and deploying shopping assistants that drive measurable conversion lift. Tell us about your store.