Professional Project
OstimGPT
A locally hosted Retrieval-Augmented Generation chatbot answering questions from university sources.
Prototype
SmartShop RAG Product Advisor
A personal AI project exploring retrieval quality for product recommendation: SQLite as the primary data source, Qdrant for vector storage, hybrid dense and sparse retrieval with BGE-M3 embeddings, reranking, product filtering, multilingual query support and evaluation datasets.
Jump to section
Overview
SmartShop is a personal project about a narrow question: what actually improves retrieval quality when the corpus is a product catalogue rather than prose?
Product queries break naive semantic search. They mix exact tokens that must match — a model number, a size — with fuzzy intent that only embeddings capture. The system combines both retrieval styles, reranks the result, and measures the outcome against evaluation datasets rather than against impressions.
Context
Recommendation systems are usually judged by anecdote: a few queries are tried, the results look reasonable, and the system is declared to work.
This project takes the opposite approach — build the evaluation datasets alongside the retrieval, so a change can be shown to help rather than assumed to.
Problem
Dense retrieval alone misses exact matches; sparse retrieval alone misses intent. Neither is sufficient for product search, and choosing between them is the wrong question.
- Exact tokens such as model numbers must be matched exactly, not approximately.
- Queries expressing intent rather than keywords must still return sensible products.
- Queries arrive in more than one language.
- Structured attributes — category, price, availability — need to constrain results, not just influence them.
- Retrieval changes need to be measurable rather than judged by inspection.
My Role
My Role
This is my own project. I built the retrieval pipeline: SQLite as the primary data source, Qdrant for vector storage, hybrid dense and sparse retrieval using BGE-M3 embeddings, a reranking stage, product filtering, multilingual query support, and the evaluation datasets used to assess retrieval changes.
Responsibilities
Retrieval
- Hybrid dense and sparse retrieval.
- BGE-M3 embeddings.
- Reranking of retrieved candidates.
- Multilingual query support.
Data
- SQLite as the primary data source.
- Qdrant vector storage.
- Product filtering over structured attributes.
Evaluation
- Evaluation datasets for assessing retrieval quality.
Technical Approach
Keep the catalogue in a relational store as the source of truth, index it for vector search, retrieve with both strategies at once, then spend the remaining effort on reranking and measurement.
- Hold the product catalogue in SQLite as the primary data source.
- Embed products with BGE-M3 and index them in Qdrant.
- Retrieve candidates using dense and sparse strategies together, so exact-token and intent-style queries are both served.
- Rerank the combined candidate set, because first-stage retrieval optimises for recall and reranking is where precision is recovered.
- Apply product filtering over structured attributes as constraints on the result set.
- Assess changes against evaluation datasets rather than by inspecting a handful of queries.
Architecture
The product catalogue is held in SQLite and embedded with BGE-M3 into Qdrant. A query runs dense and sparse retrieval in parallel; the combined candidates are reranked, then constrained by product filters over structured attributes. Evaluation datasets measure the effect of changes to any stage.
- Catalogue: products held in SQLite as the primary data source.
- Indexing: products embedded with BGE-M3 and stored in Qdrant.
- Query: a multilingual query enters the retrieval pipeline.
- Hybrid retrieval: dense and sparse strategies produce a combined candidate set.
- Reranking: candidates are reordered for precision.
- Filtering: structured product attributes constrain the final result set.
- Evaluation: datasets measure retrieval quality across changes.
Features
Hybrid retrieval
Dense and sparse strategies together, so exact tokens and intent both work.
BGE-M3 embeddings
Multilingual embeddings backing the dense side of retrieval.
Reranking
A second stage that recovers precision after a recall-oriented first pass.
Product filtering
Structured attributes constrain results rather than merely influencing them.
Evaluation datasets
Retrieval changes are measured rather than assumed to help.
Challenges
Semantic search returns plausible-looking products that are the wrong model or the wrong size, because embeddings blur exactly the tokens that matter most.
Pairing sparse retrieval with dense retrieval keeps exact tokens addressable while intent-style queries still work.
Retrieval changes are easy to convince yourself about by trying a few queries.
Building evaluation datasets alongside the pipeline made the effect of a change something to measure instead of something to argue about.
Decisions and Tradeoffs
| Decision | Alternative considered | Why |
|---|---|---|
| Keep SQLite as the primary data source with Qdrant as the vector index. | Treating the vector store as the system of record. | A relational store keeps structured attributes queryable and authoritative; the vector index stays a derived artefact that can be rebuilt. The cost is keeping the two in step. |
| Add a reranking stage. | Returning first-stage retrieval results directly. | Reranking costs latency on every query, but first-stage retrieval is tuned for recall — without a second stage, precision stays capped. |
Result
A working prototype that retrieves products using hybrid dense and sparse strategies with reranking, constrained by structured product filters and evaluated against purpose-built datasets. It is an experimental personal project, not a deployed product.
Lessons Learned
- Hybrid retrieval is not a hedge. Dense and sparse strategies fail on different queries, which is exactly why running both is better than picking one.
- Reranking is where precision is won, once first-stage retrieval has done its job on recall.
- Building the evaluation set first changes how you work on retrieval — it turns opinions into measurements.
Technology Stack
Data
AI and retrieval
Backend
Screenshots
The entry point: suggested analyses stand in for the empty state, and any of them can be typed as a free-form question instead.
An exported answer. The filters the query resolved to are shown alongside the results, and every recommendation cites the product record it came from.
Let’s discuss your project
Have a software, AI, RAG, or web application project in mind? Let’s discuss what you need and determine the right technical approach.