PRISM
Every document, refracted into a verified answer.
Prism turns papers, reports, and long documents into a queryable workspace — hybrid retrieval, grounded generation, and claim-level verification, in one pipeline.
Research workflows weren't built for how much you actually read
Papers and long documents accumulate faster than anyone can read them, and the part that matters is usually buried on a page nobody reopens.
An ordinary AI chat produces a fluent answer with no link back to a source, so every claim has to be manually re-checked before it can be trusted.
Plain keyword search returns pages that contain the words but not the meaning, which means skimming by hand for the passage that actually answers the question.
Even when an answer looks right, checking it against the original material is a separate, tedious step most workflows quietly skip.
An auditable pipeline, not a black box
Prism exposes the entire retrieval-to-generation pipeline — fused hybrid search, reranked chunks, inline citations, and per-claim verification — so every answer can be traced back to the exact passage it came from.
Everything needed to trust an AI answer
PDF, DOCX, DOC, TXT, URLs, and pasted text all flow through the same chunking and embedding pipeline, tracked through a background job.
Dense embeddings and BM25 keyword search are fused with reciprocal rank fusion, then reordered by a cross-encoder reranker.
Every sentence in an answer is scored against its retrieved evidence and labeled supported, uncertain, or unsupported.
TLDR, key concepts, methodology, results, and limitations returned as one brief instead of a wall of text.
Inspect the exact chunks, source documents, and similarity scores behind every answer, not just the final text.
Run recall@5, mean reciprocal rank, and groundedness metrics against your own workspace whenever you need a check.
One workspace, every stage of the pipeline
Eight pages, all connected — ingest a document and follow it through retrieval, generation, and verification.
Dashboard
Live metrics across documents, generations, and verifications.
Ingest
Add sources by upload, URL, or pasted text, with staged job tracking.
Library
Every ingested document, searchable and sortable by size and chunks.
Workspace
Query your research in plain language, grounded in your documents.
Source Trace
See the retrieved chunks and similarity scores behind any answer.
Verification
Review each claim, labeled supported, uncertain, or unsupported.
Evaluation
Run recall, MRR, and groundedness metrics across your workspace.
Settings
Manage your account, profile, and password from one place.
From document to verified answer, four stages
Upload a PDF, DOCX, DOC, or TXT file, fetch a URL, or paste raw text. Content is verified against its file signature, chunked, and embedded with all-MiniLM-L6-v2.
A query runs against dense vector search and BM25 keyword search in parallel, merged with reciprocal rank fusion, then reordered by a cross-encoder.
Llama 3.3 70B on Groq answers strictly from the retrieved chunks. Every fact carries an inline source marker tied to the exact chunk it came from.
The answer is split into individual claims and matched against the retrieved context, then labeled supported, uncertain, or unsupported with a confidence score.
Built for the way research actually happens
Move through papers and reports faster, with every summary traceable back to the exact passage it came from.
Ask direct questions about assigned readings and get answers that cite the source instead of a paraphrase to double-check.
Turn a folder of internal documents into a queryable workspace, with verification standing in for a manual re-read.
Cross-reference client reports and source filings without losing track of which document a figure came from.
What you actually get
Confidence scores and claim-level verification mean nothing gets cited without a traceable source.
Structured summaries surface methodology, results, and limitations without a full read-through.
Source Trace shows the exact chunks and similarity scores behind every answer.
Hybrid retrieval and reranking run in the background, so answers stay quick and evidenced.
Composed from proven, swappable pieces
Every layer of the pipeline is a distinct, independently replaceable component.
Not another chat-with-your-PDF wrapper
Your documents stay yours
- Session-scoped, signed JWT authentication on every request
- Inference provider keys never reach the browser
- Documents and chunks are isolated per account, never pooled
- No third-party tracking of ingested document content
Frequently asked questions
PDF, DOCX, DOC, and TXT uploads, URLs, and pasted raw text, all through the same ingestion pipeline.
Stop guessing whether your AI answers are true.
Ingest your first document and watch the full retrieval, generation, and verification pipeline run end to end.