Research Assistant

A Multi-Agent Research Assistant with Verified Evidence and Publication-Ready Output

Planner, reader, writer and critic agents turn a research question into an IEEE-cited literature review, typeset as LaTeX and PDF. Evidence comes from full-text papers, and every quote is checked against its source before it can be cited.

4 agentsplan, read, write and peer-review
Full textreads papers, not just abstracts
Every quoteverified against its source
.tex .bib .pdfIEEE-formatted, ready to submit

Overview—Language-model assistants write fluent literature reviews, but their evidence is unreliable: citations point to papers that say something else, and numbers drift. This system treats evidence as data to be verified. Agents plan the search, read open-access papers in full, and extract claims with verbatim quotes. A verification step discards any quote that the source doesn't contain. A critic agent reviews each draft and sends the system back to search for whatever is missing. The result is an IEEE-style document with references built from real bibliographic records. It runs as a secure multi-user web service.

IThe Problem

Ask a language model to summarize a field and you get something that reads like a review. Its references may be invented or misattributed, and its numbers can be off by a digit with no visible sign. For a researcher, checking all of that takes as long as writing the review by hand.

This project starts from one rule: the model is a strong reader and writer, but never the authority on what a source says. References come from a scholarly index, quotes are checked against the retrieved text, and the methodology section is written from the actual search log.

IIHow It Works

Each stage is a separate agent with a structured output, and each agent's model can be chosen independently, across Claude and Gemini.

  • PlannerSplits the question into sub-questions and search queries. The user can review and edit the plan before the run continues.
  • ReaderReads each source and extracts claims, each backed by a verbatim quote.
  • WriterDrafts the document with IEEE [n] citations, tables and diagrams.
  • CriticPeer-reviews the draft, flags unsupported claims and gaps, and requests follow-up searches.
Agent pipeline: question, planner, search, full text, reader, quote check, writer, critic, exports, with a critic feedback loop to search Question kind · depth Planner sub-questions · queries optional human plan review Search web · scholarly Full text open-access papers Reader claims + verbatim quotes findings Quote check against the source Writer IEEE [n] citations Critic peer review Exports .md · .tex · .bib · .pdf History runs · usage · cost verified draft approved gaps → follow-up searches revise
LLM agent deterministic stage verification gate critic feedback
Fig. 1. The research pipeline. The critic's feedback loop is what separates this from a single long prompt. Gaps and unsupported claims become new searches, and the new evidence passes the same quote check before the writer revises.

IIIVerified Evidence

Full texts, not abstracts. For each scholarly work, the system looks for an open-access full text across several sources. The reader then works from the paper's own methods and results. Paywalled works fall back to their abstract, and every source is labelled full text, abstract only or web page, so both the writer and anyone reading the review can see how much weight each one carries.

Every quote is checked. Before a finding can be cited, its quote must be found in the text the reader actually saw. The check doesn't use a language model. It forgives the noise that real documents contain, such as PDF line breaks, typography and omitted citation markers. It does not forgive a changed number, a truncated figure, or a plausible sentence the source never wrote.

Source passage (as extracted from a PDF)

Rayleigh scattering is strongly wavelength-dependent: its intensity scales with the inverse fourth power of wavelength [23]. In our measurements, the scattering coefficient at 450 nm was 0.92 of the predicted value, while at 700 nm the agree- ment fell to 0.81 (n = 1,204 samples; p < 0.001).

Table IQuotes an agent might return for this passage

QuoteWhat it testsVerdict
“the agreement fell to 0.81”Word split across a PDF line breakkept
“…wavelength. In our measurements”Citation marker [23] left outkept
“its intensity scale with the inverse…”Small transcription slipkept
“…strongly wavelength-dependent … at 450 nm was 0.92…”Quote with an elisionkept
“…at 450 nm was 0.9 of the predicted value”Truncated numberrejected
“…at 450 nm was 0.95 of the predicted value”Altered numberrejected
“n = 204 samples; p < 0.001”Digits dropped from 1,204rejected
“Rayleigh scattering explains why sunsets appear red”Plausible, but not in the sourcerejected

A literature review's methodology section reports what really happened: which databases were searched, how many works were found and read in full, and how many quotes were verified or rejected.

IVPublication-Ready Output

One question produces a research report, a literature review with a comparison table of key studies and a research-gaps section, or a background section ready to drop into a paper.

Real references

IEEE references are built from bibliographic records, never written by the model, and numbered in order of first use.

LaTeX and BibTeX

Every document comes as an IEEEtran .tex paper with its .bib, ready for Overleaf or a journal template.

Proper figures

Diagrams become vector figures in LaTeX and Mermaid in Markdown, never ASCII art.

PDF on demand

A compiled PDF comes with every run, built by a self-contained LaTeX engine in a sandbox.

VBuilt for Production

The same system runs on a laptop for one researcher, or as a public web service for many.

Accounts

Sign-up with email confirmation, secure sessions, and each user's history kept private.

Your keys, your models

Users bring their own API keys, encrypted at rest, and choose the model for each agent.

Cost control

Every run reports token usage and cost, and a per-run spending cap stops a run before it goes over.

Reliable long runs

Research runs on a durable job queue. Progress streams live to the browser, and jobs recover cleanly from crashes and deploys.

Secure by design

Protection against server-side request forgery, cross-site requests and brute-force login, plus sandboxed document builds.

Tested end to end

Every change is checked by automated tests and LaTeX builds, plus a full production stack booted in CI.

VITechnology

AIClaude and Gemini, with structured outputs
RetrievalWeb search, OpenAlex, open-access full texts
BackendPython, FastAPI, SQLAlchemy
DataPostgreSQL, Redis
Front endNext.js, React, Tailwind
TypesettingLaTeX (IEEEtran), BibTeX, TikZ
OperationsDocker, Caddy, GitHub Actions

What's Next

  • Upload your own PDFs, including papers you can access behind a paywall.
  • Ask follow-up questions about a finished report.
  • Scheduled monitoring of a topic, and detection of contradictions between sources.

See it on your own research topic

The best way to judge it is on a question you know well. I'm happy to run a live demo.