Short answer
CiteFlow's Semantic Search BETA can search by meaning after you download and verify E5 Large INT8 or BGE-M3 Full Precision. Use exact search when you know the wording. Semantic results suggest passages to review; they are not answers, proof or a measure of study quality.
When to use this method
Zotero indexes PDF text for full-text search; CiteFlow adds optional on-device meaning-based search with E5 or BGE-M3 inside a single-user project.
Before you start
Add the PDFs, make image-only pages searchable with OCR when needed and leave enough disk space for a model and its local index. Exact search looks for matching words. Semantic Search, marked BETA, compares numerical representations of meaning. An index is the local data built from the project text so later searches can run on the PC.
How we checked this guide How local retrieval, model choices and evidence limits were checked.
- What we reviewed
- Reviewed normal text search, pinned E5 Large INT8 and BGE-M3 Full Precision installation, local indexing, multilingual retrieval, null results and source reopening in CiteFlow.
- 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 normal text search first
Use CiteFlow's normal search when you know a distinctive phrase, title or identifier; it does not require a semantic model.
Choose a resource profile
Choose E5 Large INT8 for an approximately 557 MiB download and an 8 GB memory recommendation. Choose BGE-M3 for an approximately 2.13 GiB download and a 16 GB memory recommendation.
Verify and build the index
Let CiteFlow check pinned model files locally and index the extracted project text without moving PDFs to a search service.
Ask a concept-level query
Choose the relevant source scope or filters, then use a specific concept that can be judged from passage context rather than a broad topic word.
Refine instead of trusting the first list
Adjust the query or source filters when results are broad. The model-specific index stays local, but OCR and extraction mistakes inside it can still affect retrieval.
Reopen and export only checked results
Check population, method, qualifier and source identity on every promising page. Export a search report or matching original pages only after that review; an export preserves the selection, not its truth.
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.
- Inspect an unrelated control query.
- 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.