Skip to main content
If you are building a chatbot, a search feature, or any assistant that answers questions from your documents, its answers can only be as good as the passages it finds. That finding step, called retrieval, depends on choices you make when building the pipeline: the embedding model, how your documents are chunked into passages, whether you add hybrid search (keyword search next to vector search), and whether a reranker re-scores the results. Truvec lets you test those choices, and combinations of them, on your own data instead of guessing.

What Truvec does

You bring a sample of your documents. Truvec splits them into passages, builds a set of test questions whose correct passages are known, and checks how often a retrieval setup finds those passages. Change the model, add hybrid search or a reranker, or try another chunking, and run the same questions again: you get clear scores, the cost and speed of each setup, and a question-by-question view of what each change fixed or broke.

Measure retrieval quality

See how often the right passage is found, and how high it is ranked, on questions about your own documents.

Combine and compare

Run embedding models with or without hybrid search and rerankers on the same questions. Keep a baseline and see what each change is worth.

Understand every miss

Open any test question and see which passages were retrieved, which expected ones were missed, and how hybrid search or a reranker moved them.

Know what it will cost

Get the cost to index your documents and to answer every thousand queries, reranking included, with the latency of each step.

What you can test

Models, hybrid search and rerankers can be mixed freely in one comparison, for example a cheap model with a reranker against a bigger model without one. Chunking is compared with one golden set per chunking, because a golden set points at specific chunks.

Who it is for

Truvec is for anyone who needs to build, tune or validate the retrieval step of a retrieval-augmented generation (RAG) system, including people who have never run an evaluation before:
  • Product and engineering teams choosing a pipeline before building, or checking whether a cheaper model, with or without a reranker, is good enough.
  • Teams with a RAG system in production who want evidence before switching models, adding a reranker or hybrid search, or changing how documents are split.
  • Anyone new to RAG who wants to understand why their assistant sometimes answers from the wrong passage.
New to embeddings or RAG? Start with How RAG retrieval works. It explains the ideas you need in a few minutes, without math.

How it works

1

Create a project and add documents

Upload a representative sample of your documents (PDF, Word, Markdown, text, CSV or JSON), or chunks you already have. Truvec splits them into passages called chunks, which you can open and read.
2

Build a golden dataset

Generate test questions with AI, or write your own. Each question is linked to the chunks that contain its answer. Review them, then lock the dataset.
3

Evaluate and compare

Run one configuration, or compare several: different models, or one baseline with hybrid search, a reranker or another model. Truvec asks every question and checks where the correct chunks appear in the results.
4

Read the results and decide

Compare quality, cost, speed and storage, then inspect individual questions to understand where each setup succeeds, where it fails, and what each change fixed or broke.

Why not just use the defaults?

Public leaderboards, chunking rules of thumb and “always add a reranker” advice are based on generic data. Your documents have their own vocabulary, structure and question types: internal product names, legal clauses, support tickets, code. A model that ranks first on a leaderboard can be beaten on your data by one that costs a fraction of the price; a reranker can lift a cheap model above an expensive one, or add cost for no gain; hybrid search can rescue questions about product codes and hurt conversational ones. The only reliable way to know is to test on your own content, which is what Truvec is for.

Quickstart

Run your first comparison in about 15 minutes.

Hybrid search and reranking

What each strategy changes, and what it costs.