Short answer
Use normal search when you know the words. Use Semantic Search BETA when you know the idea but not the wording. Install one local model, build the project index, and open every suggested passage before you use it.
When to use this method
Use this method when different papers may describe the same idea with different words. It searches PDFs already in your project.
Before you start
Add the PDFs. Apply OCR to image-only pages. Leave enough disk space for one model and its index. Exact search finds matching words. Semantic Search BETA finds related meaning. An index is a local map of the project text used for later searches.
Technical note Confirmed behavior and an important limit.
- What we confirmed
- PDF extraction, embeddings and queries run on the PC after the optional model download. Embeddings are numerical representations used to compare meaning. Similarity scores identify related passages but do not assess truth, study quality or applicability.
- Important limit
- A high-ranking passage about the same intervention can use a different population. Read the surrounding method and limitation before treating it as evidence for the review question.
Install, index and verify a local model
Try exact search first
Search for a known phrase, title, or identifier. You should see direct word matches without installing a semantic model.
Choose one model
Choose E5 Large INT8 for the smaller download and lower memory use. Choose BGE-M3 for broader multilingual retrieval on a computer with more memory. The detailed sizes and memory guidance appear below.
Install and build the index
Let CiteFlow verify the model and index the project text. You should see the component and index marked ready before searching.
Search for one clear idea
Choose the source scope and enter one specific idea. You should see passages with source names and pages that you can open.
Try an unrelated control query
A control query is a test phrase that should not match your research topic. Search for one clearly unrelated idea. If the same passages still rank highly, treat the search results as unreliable and inspect the index and source text.
Fix an incomplete index
If the component is missing or failed, use Manage to install or repair it. Apply OCR when scanned pages are marked partial. Rebuild a stale index after source changes. Restart an index that was canceled.
Open and check each result
Open a promising result. You should see the original page and surrounding text. Check its population, method, qualifiers, and source before keeping or exporting it.
Model size and memory details
E5 Large INT8 is an approximately 557 MiB download and has an 8 GB memory recommendation. BGE-M3 Full Precision is an approximately 2.13 GiB download and has a 16 GB memory recommendation. Each model builds its own local project index.
Semantic-search limits
- Installing a large model for an exact identifier search
- Treating rank as study quality
- Ignoring OCR or extraction errors in indexed text
- Citing a match without reopening its page
Checkpoint: Retrieval checks
- Compare one normal text query with one conceptual query.
- Confirm the computer has enough storage and memory for the selected model.
- Run an unrelated test phrase and check that irrelevant results do not rank highly.
- Open the original page for each retained match.
- Confirm exported matches use the intended source scope.
CiteFlow
Build a local E5 or BGE-M3 index, refine meaning-based retrieval and reopen every result before using it as evidence.