Trace a concept across works and languages, compare how authors use it, and attach your notes to specific passages.
THE RESEARCH WORKFLOW
Import source material and extract text or transcribe audio. Segment it into records that retain a location appropriate to the medium, such as a printed page or an audio time range.
The source, the extraction method, and the location behind each record. The search index can be rebuilt from the corpus.
Use a metadata schema to define which fields matter and what evidence they require. Review proposed values and their supporting passages. Confirm or correct the values before using them as reviewed metadata.
Each field assertion can carry its evidence, provenance, and review state. A reviewer’s decision remains distinct from the model’s proposal.
Choose a corpus and retrieval settings, or supply passages yourself. DerridAI assembles the selected passages and their source metadata into an evidence packet for the model to use in its synthesis.
The passages supplied to the model and the run’s retrieval settings. Attribution fields help distinguish the speaker from the person whose position is being discussed.
Follow a citation back to its passage and check whether it supports the claim, including its attribution and qualifiers. Compare runs using the same question to see how retrieval settings or model choice affect quality and latency.
The answer, available claim-to-evidence links, and run diagnostics. The researcher decides whether the interpretation holds up.
A quotation can introduce a position the speaker goes on to challenge. DerridAI has separate fields for who is speaking, whose view they are discussing, and how they respond to it.
Ingest documents, text, images, and audio with format-aware extraction. Segment material into records while preserving source identity, page or time coordinates, and extraction provenance.
Explore the corpus workflow
Define fields, evidence requirements, and review policies. Add guidance for a particular run. Reuse evidence-bound reviewed precedents and corrections to inform later proposals.
Combine semantic and lexical retrieval, MMR, deduplication, and optional reranking. Or bypass retrieval and synthesize from passages you choose. Inspect the evidence packet alongside the response.
Evaluate relevance, traceability, attribution, evidence fidelity, and conceptual precision. Inspect pipeline timings and retrieval diagnostics. Re-run a question with different parameters to investigate quality and latency.
Explore Research in the user guide
DEPLOYMENT AND GOVERNANCE
WCAG 2.2 AA is a release criterion for the application. Its interface is designed for keyboard use, visible focus, reflow, reduced motion, and forced colors. Automated checks also cover the reference interface for exported sites.
Read about export accessibilityDerridAI ships with English and French. Administrators can add any interface language by installing its translation dictionary. Interface, source, and answer languages are separate, so quotations can retain their original wording.
See how cELF handles languageRun the application with Docker Compose and use it in a modern browser. Connect a local Ollama instance or an OpenAI-compatible endpoint. Exported research sites can be opened as a standalone HTML file or served with the nginx bundle.
View installation requirementsConfigure metadata fields, evidence requirements, provider profiles, and versioned pipelines. The browser SDK supports custom publication interfaces, and a read-only GraphQL API exposes typed cELF queries.
Read the architectureSelf-host with a local model to keep corpus storage and inference on infrastructure you control. If you configure a remote provider, it can receive prompts and selected evidence. The application enforces roles at the API boundary, including restrictions on corpus access and export.
See how access is controlledDerridAI lets you work with copyrighted sources locally. With a local model, you can search and analyze them without sending them to a provider or publishing the corpus. Keeping the source files local and sharing citations or permitted material helps avoid unauthorized redistribution. Check what rights or exceptions apply before using or sharing material.
Read the cELF rules for restricted sourcesCapta-Enriched Locatable Factum
cELF defines a durable research record that can be traced to its documentary source. DerridAI is the originating reference implementation.
Explore the cELF specificationKeep the source identity and a precise locator such as a page, a text range, an image region, or an audio time span.
Give records stable identities and explicit revisions. Attach field assertions with their evidence, provenance, and review authority. Treat search indexes and embeddings as rebuildable views of the records.
Connect a generated proposition to the exact evidence and record revision used to support it. This lets a reader verify both the citation and what the passage actually establishes.
Bring your own PDFs, plain text, scans, or audio recordings. DerridAI can extract text from images and transcribe audio.
A local Ollama setup works without a hosted model account or API key. If you connect to a hosted model, use the credentials that service requires.
Open the cited passage to see whether it supports the claim, whether the answer attributes the view to the right person, and whether the surrounding text changes its meaning.
Readers can open a standalone HTML export in their browser. They can browse and search the works you chose to include.
The repository includes installation instructions for Docker Compose.