01.1 Documents
PDF, Office, Markdown, HTML, text, and code.
Upload text-based PDFs, Word documents, presentations, spreadsheets, CSV or TSV data, Markdown, HTML, logs, and common source-code formats.
LIBRE / PRODUCT Retrieval with evidence CURRENT CAPABILITIES
Knowledge
Know where the answer came from.
Libre turns the documents you can access into usable chat context without severing the answer from its source. Exact terms remain searchable, semantic matches add reach, and cited locations stay visible in the conversation.
- Keyword search by default
- BM25
- Vector + lexical when enabled
- Hybrid
- Page, slide, sheet, section
- Source-level
01 Ingestion
Useful formats keep their place in the source.
A document is more useful when an excerpt can still say where it lived. Libre carries structured provenance through extraction and avoids reprocessing unchanged uploads in the same scope.
01.2 Provenance
Keep the page, slide, sheet, or section.
Extraction records source segments so retrieved excerpts and the Sources rail can identify the location a claim came from—not just the filename.
01.3 Bounded processing
Parse Office files without a native document runtime.
DOCX, PPTX, and XLSX are unpacked through a bounded in-repository parser. Identical bytes uploaded again to the same scope are deduplicated.
Go deeper
02 Retrieval
Find the identifier and the idea.
Keyword and semantic retrieval solve different misses. Libre keeps the lexical path available at all times, then fuses it with vector ranking when embeddings are configured.
02.1 Keyword
BM25 is always available.
Exact identifiers, names, and rare terms can rank without an embedding model. If embedding generation fails or is disabled, document search continues lexically.
02.2 Hybrid
Fuse vector and lexical rankings.
With embeddings enabled, reciprocal-rank fusion combines semantic similarity with BM25 so a vague semantic neighbor does not automatically beat a strong exact match.
02.3 Encrypted content
No plaintext full-text index beside ciphertext.
Lexical scoring runs in process over decrypted chunks the requester may access. Libre deliberately avoids persisting a plaintext token index for encrypted document chunks.
Go deeper
03 Answers
Use the right amount of source context.
Retrieval is the efficient default. When the task genuinely depends on the entire file, a conversation can opt into full-document context instead.
03.1 Citations
Carry source locations into the reply.
Retrieved excerpts include the source name, chunk, score, and structured location. The Sources rail groups the cited locations under their documents.
03.2 Full-document mode
Read the whole extracted document when it fits.
A chat can send complete extracted content instead of selected passages. A configurable token guard falls back to retrieval when the attached material is too large.
03.3 Knowledge tools
Let the model search and read during a turn.
Built-in tools can list in-scope documents, search cited passages, and read bounded windows so the model can investigate a file iteratively rather than receiving everything up front.
See the tool loop