LLM Book Editor - Local Manuscript Reviewer

Description

I built a local LLM manuscript reviewer to give authors practical editing guidance without sending drafts to external services. The goal was to scan full length manuscripts, keep context intact, and return actionable feedback that feels like a real editorial pass. I designed it as a private, on device workflow because authors care about ownership and confidentiality. I also wanted it to fit into a real revision cycle, where feedback is applied in rounds rather than a single prompt.

To handle long documents, I used context aware chunking and a local embedding index so the model could see local passages while still respecting global themes. The pipeline splits chapters into structured segments, indexes them, and retrieves relevant context for each editorial pass. I built separate passes for structure, clarity, and line level edits so the output stays focused and avoids repetition. The core workflow is in C#, with TorchSharp powering embeddings, SQLite storing intermediate artifacts, and Docker keeping runs consistent. This kept the workflow fast, repeatable, and fully local.

I kept the output format practical. Each pass produces summaries and concrete suggestions rather than long essays, which makes it easier to apply changes. Settings are explicit and adjustable, so I can tune tone, strictness, and scope without altering the code. I added simple checks to avoid conflicting feedback and to surface missing context. The pipeline is deterministic enough to compare revisions across versions, which helps authors track progress without reprocessing everything from scratch. It also supports incremental runs, so I can re-review a single chapter or section quickly.

The result is a repeatable review loop that reduces manual editing while keeping the author in control. It gives me a clear second set of eyes that catches issues across a long manuscript and helps maintain voice and consistency. It also creates a shared reference point for discussions with other editors or reviewers. If I extend it further, I would add custom style guides, domain specific checks, and deeper cross chapter consistency scans. This project shows how LLMs can be useful when the workflow is designed around real editing needs and privacy first constraints.