Weighted vector search with Superlinked

Superlinked builds vector compute for search that has to understand more than one thing at once. We built a set of retrieval applications on it, each one a case where a single embedding was not enough and the search had to blend meaning with field weights, recency, fuzzy matching, or domain-specific ranking.

The common thread is a limit of plain vector search. It squashes a whole record into one embedding and compares it with one query embedding. That works when similarity is the only signal, and falls apart when similarity is one signal among several.

When One Embedding Flattens the Question

Think about what people actually ask. A game-store query mixes genres with free text. A property search weighs price ranges and location constraints alongside a description. A research feed needs recency to count as much as topic, and a fintech decision wants a model’s own prediction to sit inside the ranking.

Each of those needs a retrieval layer where fields can be weighted, near-matches can be scored, and structured logic can run next to semantic similarity instead of being flattened into it. Superlinked’s weighted vector spaces gave us that, so we used them to add a different kind of signal in each project.

Four Searches, Four Different Signals

The Steam recommender pairs Superlinked with LlamaIndex. It resolves multi-field queries, like strategy co-op games with sci-fi elements, through weighted vector spaces rather than one text embedding, so genre and description each pull their own weight.

The real-estate natural-language query agent puts strong weighting on price and applies fuzzy matching across property constraints. People describe homes loosely, so imprecise phrasing degrades gracefully instead of failing on exact tokens.

The research-paper agent keeps text relevance and recency in separate spaces and balances them, then layers GPT-4 summarization on top. A reader gets recent work that is also on topic.

Putting a Prediction Inside the Ranking

The fintech agent combines an ML loan-prediction model, at 84% accuracy, with semantic insurance-claim validation over LanceDB. Claims are matched on meaning rather than exact fields, because the same situation gets written up in many different ways. Those claim-matching workflows reached approval rates above 50%.

• A Steam recommender handling multi-field queries such as strategy co-op games with sci-fi elements.
• A real-estate NLQ agent with strong price weighting and fuzzy matching across property constraints.
• A research-paper agent balancing text relevance and recency, with GPT-4 summarization.
• A fintech agent pairing 84% accurate loan prediction with semantic insurance-claim validation.
• Insurance-claim matching with more than 50% approval rates through semantic validation.

Read the Builds

Each project has its own write-up: the Steam game recommender with custom retriever, the real estate natural-language query agent, the research paper agent, and the fintech agent.

Across them we worked with Superlinked, LlamaIndex, LanceDB, GPT-4, GPT-3.5 Turbo, Mixtral, Sentence Transformers, Pandas, PyArrow, scikit-learn, and Python.