Grafy News
Try Grafy free

MeReader Offers Offline, Privacy‑First AI Companion for Narrative eBooks

A new research paper introduces MeReader, an offline, progress‑aware AI assistant that protects privacy and avoids spoilers while helping readers understand.

Illustration for the news story: MeReader Offers Offline, Privacy‑First AI Companion for Narrative eBooksSource: Grafy Research Lab

MeReader, a newly published research prototype, brings an offline, privacy‑preserving AI assistant to narrative eBook reading, aiming to keep the reading experience spoiler‑free while still offering contextual help. The system, described in a preprint from the Grafy Research Lab, targets consumer hardware and positions itself as an alternative to cloud‑based LLM services that process whole books at once.

What is MeReader and why it matters?

MeReader is an on‑device AI companion that activates only around the reader’s current location in a book. Unlike popular services such as OpenAI’s ChatGPT or Meta’s LLaMA that require uploading an entire document to a server, MeReader runs locally, ensuring that the text never leaves the user’s device. According to the paper, this design directly addresses two common complaints: the risk of narrative spoilers leaking from full‑text processing, and the privacy concerns of sending copyrighted or personal reading material to external servers.

How does MeReader keep reading offline and private?

The authors built MeReader to operate without any internet connection after the initial installation. All models, embeddings, and index structures are stored on the device. By confining both retrieval and generation to the local environment, the system eliminates any network traffic that could expose the user’s reading habits. The paper notes that the approach leverages “consumer hardware,” meaning it can run on typical laptops, tablets, or e‑readers that have modest CPU/GPU resources. Model sizes in the evaluation range from 2 billion to 5 billion parameters, which the authors found sufficient for high‑quality responses when optimized for on‑device inference.

What retrieval techniques does MeReader combine?

MeReader uses a hybrid retrieval pipeline that blends three classic methods:

  1. Vector similarity – dense embeddings of text chunks are stored in a local vector database, enabling semantic matching.
  2. BM25 – a proven term‑frequency based ranking that captures exact keyword matches.
  3. Summary embeddings – lightweight representations of chapter‑level summaries that help the system locate broader context without scanning the entire book.

By integrating these signals, the system can retrieve passages that are both semantically relevant and lexically precise, while staying within the user’s current reading window. The paper highlights that this combination improves “fidelity to the reader’s current context” compared with using any single method alone.

How does progress‑aware gating prevent spoilers?

A core innovation of MeReader is its progress‑aware boundary mechanism. The assistant is aware of the reader’s exact position – measured in characters or paragraphs – and restricts any query or generation to content that appears before that point. In practice, if a user asks about a character’s fate that is revealed later in the book, the system will either refuse to answer or provide a safe, non‑spoil response. Experiments in the paper show that this gating effectively eliminates future‑content leakage without noticeably degrading the relevance of retrieved passages.

How does MeReader perform compared to cloud‑based LLM assistants?

The researchers evaluated MeReader on a set of classic English novels, using both quantitative comprehension metrics (such as question‑answer accuracy) and qualitative judgments from independent LLM reviewers. They discovered a “quality‑latency frontier”: mid‑sized local models (2 b–4 b parameters) achieved comprehension scores comparable to larger cloud models (5 b+ parameters) while delivering faster response times on typical hardware. Importantly, the offline setup avoided the privacy and spoiler issues inherent to cloud services. The authors conclude that a well‑tuned local model can provide a “good alternative to cloud‑dependent services” for narrative reading.

Using Grafy to explore similar AI reading tools

While MeReader itself is not yet a commercial product, Grafy’s Reader app offers a comparable playground for researchers and creators interested in AI‑augmented document interaction. Reader lets you upload PDFs or eBooks, ask natural‑language questions, and receive AI‑generated answers, all within the Grafy platform. Because Reader runs on Grafy’s multimodal backend, you can experiment with different local models, adjust retrieval settings, and even try a progress‑aware prompt strategy by limiting the visible text range. Pairing Reader with Chat (which provides side‑by‑side model comparison) lets you see how a cloud model’s answer differs from an on‑device one, highlighting the privacy‑performance trade‑offs that MeReader aims to resolve.

For developers who want to prototype a MeReader‑style workflow, the combination of Reader and Chat offers a low‑friction path: import a novel in EPUB or PDF format, slice it into sliding windows that respect the current reading position, and connect those windows to a local LLM via the Grafy Code studio. This mirrors the hybrid retrieval stack described in the paper—vector search can be built with Qdrant, BM25 with SQLite, and summary embeddings with a lightweight encoder—all of which are supported by Grafy’s cloud‑agnostic model zoo.


Source: https://arxiv.org/abs/2611.18513

Frequently asked questions

Does MeReader require an internet connection to answer questions?

No. After the initial installation, MeReader runs entirely on the user’s device, using locally stored models and indexes, so no network traffic is needed for retrieval or generation.

How does MeReader avoid revealing future plot points?

The system tracks the reader’s exact location in the text and blocks any query that would require information beyond that point, either refusing to answer or providing a generic, non‑spoiling response.

Can Grafy’s Reader app replicate MeReader’s offline capabilities?

Reader can be configured to use local models and custom retrieval pipelines, allowing users to experiment with offline, privacy‑focused question answering similar to MeReader, though it does not include built‑in progress‑aware gating out of the box.


Source: Grafy Research Lab — MeReader: An Offline, Privacy-Preserving, Progress-Aware AI Assistant for Narrative eBook Reading