Libriq

Organizations accumulate knowledge in books, manuals, standards and reports that almost nobody can consult at the moment they need it. Keyword search returns files, not answers.

A general AI assistant answers fluently but does not say where the information came from. Where the source matters, an answer without an origin is no use.

Libriq takes books and documents in ten formats and turns them into a searchable knowledge base. From there, a person can:

• talk to the collection and receive answers that cite the source passage;
• run semantic search, which finds passages by meaning and not only by words;
• create courses, with lessons written from the books and a plan shown before the course is created;
• generate quizzes and flashcards from the content;
• track their own study progress on a mastery map by topic.

The ten formats are EPUB, PDF, DOCX, MOBI/AZW3, TXT, Markdown, HTML, ODT, RTF and FB2. Libriq indexes each file automatically once it is uploaded, and the book is ready to be questioned as soon as indexing ends.

Each document is split into passages, and each passage is indexed by its meaning in a vector database. When someone asks a question, Libriq first retrieves the most relevant passages. Only then does it ask a language model to write the answer from them.

That is why every answer points to the passage that supports it. Each citation shows the book, the chapter and the passage itself, with a link that opens the book at that point, so the reader can check.

Retrieval itself happens in two stages. A vector search first gathers candidate passages by similarity of meaning. A reranking model then reads each candidate in full and keeps only the five most relevant for the answer.

Each citation also carries a relevance score, shown as a percentage, so the reader can see how closely a passage matches the question before opening it. The language model, the embedding model and the reranking model are each pluggable: any of the three can be replaced by another, so the platform is not tied to a single provider’s models.

This way of answering, which retrieves passages and only then writes, is known as RAG. Alongside it, Libriq keeps a concept graph of the collection. Concepts are extracted from the books and linked by relations: for example, one concept requires another, is part of another or uses another. The graph shows what the collection supports on a topic and what is best learned first.

Libriq builds courses from that graph. Whoever creates a course gives it a title, and before anything is created the platform shows the lesson plan the collection supports. They can use that plan alone, extend it to the number of lessons they want with concepts from outside the collection, which are marked as such, or generate the course freely. Inside each lesson, a paragraph that cites no source from the collection is marked, so whoever approves the lesson knows what to check.

Libriq also exposes a REST API, protected by API keys, and an MCP server. Through them, other systems use the collection as a source of knowledge.

Libriq serves teams that need to consult and teach a large body of knowledge: people who train others, people who work with standards and contracts, and people who want to give an assistant a knowledge base it can rely on. In every case the value is in the same place: the answer arrives with its source.

Books are grouped into collections inside a team, so each team works on its own library. A course generated from that library can stay private, be shared with the whole team or be shared with chosen people, and each person’s progress through it is recorded separately.

Libriq is in production. It is also the knowledge retrieval layer of other systems built by Sigil Labs: a public policy copilot that answers by citing legislation, and a corporate legal assistant that answers only from an approved document library.