RAG · Document Analysis
RepeatRadar
Semantic Repetition Detection for Long Documents
Finding semantically repeated ideas across a 100+ page document, not just repeated text.
Problem
While writing a PhD thesis, the same idea, definition, or argument can end up restated in more than one chapter without the author noticing: re-reading the whole document to catch it does not scale. Exact-text search does not help either, since the repeated idea is usually reworded, not copy-pasted, so finding it requires comparing chapters on meaning rather than on matching strings.
Pipeline
Chapters are chunked and embedded with Gemini embeddings, then indexed in FAISS for nearest-neighbour search. Chunk pairs whose similarity exceeds a fixed threshold are flagged as candidate repetitions and passed on for human review.
Chunk, embed, search, review
Example
Illustrative repetition example
Illustrative repetition example
Chapter 2
The evaluation protocol separates retrieval quality from final answer quality so that improvements can be attributed to the correct stage of the pipeline.
Chapter 4
Retrieval and generation are assessed independently, allowing changes in answer quality to be traced back to the stage that produced them.
Suggested fix
Review decision
Illustrative passages, not real thesis text. No similarity score is shown; RepeatRadar flags candidates qualitatively above a fixed threshold.
Review
Suggested action
- · Rewrite
- · Shorten
- · Cross-reference
Review decision
- · Accept
- · Reject
- · Edit
RepeatRadar does not silently rewrite the document. It proposes a candidate and leaves the final editorial decision with the author.
Integration
The pipeline is served through FastAPI, with a Streamlit interface for reviewing flagged passage pairs across a document. The same repetition-check is also exposed over MCP, so it can be invoked as a tool from an MCP-compatible client instead of only through the web UI.
Limitations
- · Flags candidate repetitions for human review rather than auto-rewriting: automatically rewriting already-reviewed thesis text would risk silently changing its meaning.
- · Similarity threshold is tuned manually per document rather than learned.