THE ARCHIVE
Aletheia reads the papers you give it and answers questions across them. Every citation in an answer is checked against the passages that were actually retrieved — so a reference either resolves to real text in your library, or the answer does not make it.
THE PROBLEM
That is the whole difficulty. A language model asked about a literature will produce plausible author names, plausible years and a plausible page number, and nothing about the output signals which parts came from a document and which were assembled to fit the sentence. Checking is slower than asking, so it usually does not happen. Aletheia is built the other way around: the checking is not optional and not the reader’s job.
WHAT IT GUARANTEES
01
Evidence passages are handed to the model with opaque IDs, and the answer is re-checked against them server-side. An ID the model invented is not a footnote that happens to be wrong — it fails validation, and the answer does not ship with it.
02
When the retrieved passages do not settle a question, the system says so instead of assembling a fluent paragraph out of the nearest available text. Most of the damage done by research assistants is done confidently.
03
Embeddings and reranking are local ONNX models — no key, no per-query spend, and no page of your unpublished work sent anywhere to be indexed. Only the final answering step calls a hosted model.
04
Held-out sets are split by question ID, so passages from one question cannot appear on both sides. It is the difference between a metric and a number that looks like one.
THE PATH A QUESTION TAKES
THE WORKSPACE
Semantic retrieval over every page you have ingested, reranked by a cross-encoder rather than by keyword overlap.
OPEN →A grounded answer with inline citations. Every cited passage is one the retriever actually returned.
OPEN →A bounded planner that breaks a question into sub-questions, answers each from the corpus, and returns the full trace.
OPEN →Where several papers agree, disagree, or simply do not address the same thing — with the passages that decide it.
OPEN →Take a claim, find what the corpus says about it, and label it supported, contradicted or unaddressed.
OPEN →What a paper actually gives you: code, data, environment, hardware — and what it leaves out.
OPEN →The three models trained for this project, their learning curves, held-out metrics, and downloadable notebooks.
OPEN →Upload PDFs and watch them parse into pages, sections, figures, tables and equation candidates.
OPEN →TRAINED FOR THIS PROJECT
Each was trained from the checked-in benchmark, split by question ID, and is reproducible from a notebook in colab/. The stance classifier sits below its own majority-class baseline. It is shown here for the same reason it is shown in the app: removing it would make the other two look like the whole story.
Multi-source relevance
89.87%
held-out accuracy, against an 84.96% lexical baseline
Neural reranker
0.717 MRR
Recall@6 0.799, reproduced exactly on a full re-ingest
Scientific stance
54%
experimental — below its majority-class baseline, and shipped saying so
Upload a PDF, wait for it to parse, and ask a question it can only answer by reading. The interesting result is the one where it tells you the corpus does not cover it.
Create an account