CNRS·Université Côte d'Azur·Institut de Chimie de Nice·Centre Inria de l'Université Côte d'Azur·Interdisciplinary Institute for Artificial Intelligence (3iA) Côte d'Azur
HolobiomicsLab

The path from a paper to an answer

Retrieval you can follow

Perspicacité connects literature discovery, content acquisition, a local knowledge base, and several ways to reason over the retrieved evidence.

Content acquisition

Find the best available version of a paper.

Start with a literature search, a bibliography, DOIs, a Zotero collection or local documents. The optional SciLEx integration brings together searches across Semantic Scholar, OpenAlex, PubMed, arXiv, HAL and DBLP. Papers retain their source-database metadata as they enter the collection.

Each paper then goes through a quality-priority routing chain, which the project documents as follows:

  1. Discovery: OpenAlex and Unpaywall supply identifiers, open-access status and the abstract.
  2. Structured full text: PMC JATS XML, then Europe PMC, then arXiv HTML, with sections and references.
  3. PDF full text: open-access PDFs, arXiv, Unpaywall, publisher APIs, or institutional access by replaying browser cookies that you provide.
  4. Abstract only: when no full text is available, the paper is served from its metadata and flagged as abstract-only.
  5. Discard: a failure is returned when nothing usable can be retrieved.

PDF text is extracted with PyMuPDF. Optional layout extraction with docling, for structured tables and figures, is off by default. Content quality and access vary, so the collection preserves the distinction between full text and an abstract.

From literature sources to a cited answerSix stages source papers, acquire text, extract passages, index a local knowledge base, retrieve evidence and produce a cited answer.AN EVIDENCE PATH YOU CAN INSPECT01SourceAcademic APIsBibTeX · local filesZotero collections02AcquireStructured full textPDF when availableAbstract fallback03ExtractParse sectionsChunk by contentPreserve metadata04IndexEmbed passagesLocal vector storeBuild knowledge base05RetrieveHybrid searchRerank passagesGather context06AnswerGenerate responseAttach citationsKeep retrieval trail

Figure The working library and paper-level retrieval metadata stay local; acquisition and configured services can make network requests.

Knowledge bases and hybrid search

Keep the collection. Retrieve the passages.

A named research collection

A knowledge base groups paper metadata and content around a topic. SQLite holds metadata and conversation records; ChromaDB stores the indexed embeddings; cached PDFs and optional capsule artifacts live on disk.

Two complementary signals

Hybrid retrieval combines BM25 keyword matching with vector similarity, so exact terminology and semantically related passages can both contribute before the selected mode produces an answer.

More than one collection

Select a knowledge base explicitly, or use automatic routing (kb_name: "auto") to score all collections against the question and query the best matches. Each collection is tied to the embedding model used to build it.

Optional paper capsules bring figure captions, code snippets, references and supplementary content into the same research workflow. They enrich the material available for retrieval when those artifacts can be acquired and extracted.

Knowledge-base concepts ↗

Six ways to ask

Choose the depth your question needs.

Begin with Basic or Advanced. More involved modes add retrieval, analysis or discovery steps, and may take more time and model calls.

Basic · a focused answer

One hybrid BM25 and vector retrieval pass against the selected knowledge base, then a single synthesis call. Use it for a specific question in a well-curated collection, or to check that your first knowledge base covers the subject you expect.

Advanced · a broader search

Expands the question into several queries, merges their rankings with weighted reciprocal-rank fusion and re-ranks before synthesis. Use it when one phrasing misses useful material or a broad collection needs a more focused evidence selection.

Deep Research · an iterative investigation

Repeats planning, retrieval, self-evaluation and synthesis over several cycles, with optional web search. Earlier guides call it Profound. It suits questions with several connected aspects that one retrieval pass is unlikely to cover.

Agentic · discovery beyond the KB

A tool-calling agent searches the knowledge base, retrieves paper content and, when SciLEx is available, searches academic databases. Use it when the answer may need papers outside the current collection. It works best with direct provider APIs; its tool use makes runtime and cost more variable.

Literature Survey · a field overview

Runs a broad search, clusters papers into themes and produces a structured survey, with checkpoint and resume support. Use it to explore a research area and identify themes for closer reading.

Contradiction · compare the claims

Retrieves relevant papers, extracts claims and groups them into agreement, disagreement and open questions. Use it to examine conflicting findings and locate the papers behind them; the groupings are a starting point for reading, not a verdict.

The command-line query command offers Basic, Advanced, Deep Research and Contradiction. The web interface and the REST API also expose Agentic and Literature Survey. Use deep_research in new calls; profound remains accepted as a deprecated alias.

Mode selection and configuration ↗

From a result to its sources

Preserve the retrieval trail.

Answers linked to a conversation can keep a retrieval record in SQLite: paper identifiers, titles, retrieval scores, knowledge bases, content types and workflow steps. Separate model-call logs hold the prompts, responses and timing recorded for inference.

A conversation can be exported as Markdown or as an RO-Crate 1.1 ZIP bundle. Knowledge bases can be exported as BibTeX with cached PDFs, or as an Obsidian vault; optional Zotero integration can push records and attachments to a Zotero library.

What the provenance record contains ↗

Follow the literature outward

Grow a collection through citation links.

Citation-graph expansion follows one hop from each seed paper, to the papers it cites and the papers that cite it. OpenAlex supplies the primary graph, with a Semantic Scholar fallback for arXiv-seeded papers, because OpenAlex under-reports preprint citations. Candidates are merged, deduplicated by DOI, then filtered and screened before ingestion.

Use this after an initial search to find related work that your keywords may not have surfaced. Results depend on source coverage and the filters you choose; a citation walk is a way to expand a reading list, not a guarantee of an exhaustive review.

Citation-graph expansion in detail ↗

Local-first in practice

Know what stays and what travels.

Stored on your machine

Knowledge-base metadata, vector indexes, cached documents, capsule artifacts, conversation history and retrieval provenance live in your local installation. The example configuration uses local sentence-transformer embeddings.

Sent when a service is used

Academic search sends queries and identifiers to source services; content acquisition contacts repositories and publishers. Model inference sends prompts and selected context to the configured provider. Hosted embeddings and optional Zotero transfers send their respective inputs or files.

The project states that it collects no usage analytics. You control the configured providers, the model route and the integrations. Ollama is an option for local inference; installing dependencies, downloading models and discovering new literature still require the corresponding network access.

Inspect the example configuration ↗