Fast take: zra-mcp (PyPI release, 13 July 2026) turns a Zotero library into a local RAG pipeline to run AI-enabled literature review on your machine. zra-mcp bundles PDF chunking, CJK-aware chunking, bilingual query expansion, ONNX INT8 embeddings (~347MB), and 36 MCP tools for discovery, reading, and review, see PyPI. If your team needs to move from retrieval to reproducible thematic insight, combine zra-mcp’s retrieval and context blocks with Evidano to shorten the path from papers to decisions.
Key Takeaways
Evidano is an AI-powered qualitative data analysis platform that ingests zra-mcp’s JSON and Markdown context_blocks to run thematic, reproducible synthesis.
zra-mcp turns a Zotero library into a local RAG pipeline that extracts PDF text, applies 52 cleaning rules, performs CJK-aware bilingual chunking, embeds chunks with bge-m3 quantized to ONNX INT8 (~347MB), and returns LLM-ready evidence blocks.
- Local-first retrieval: zra-mcp runs on your machine with Zotero local API integration and produces JSONL retrieval traces for auditability.
- Bilingual support: zra-mcp includes Chinese↔English expansion and a built-in CN↔EN dictionary (300+ pairs) plus OPUS-MT on-demand translation.
- LLM-ready outputs: zra-mcp returns dual-format JSON metadata and Markdown context_blocks (blockquote evidence with relevance tiers) optimized for prompt consumption.
- Fast prototyping: the MPO toolset includes 36 MCP tools and a 60-query evaluation suite to benchmark recall and ranking.
What to know at a glance (source)
zra-mcp optimizes retrieval precision for academic PDFs on consumer hardware and returns LLM-friendly evidence blocks (blockquote plus relevance tiers).
The project page and README document the pipeline and tools: PyPI (accessed 13 July 2026).
- Local-first RAG: runs entirely on your machine (Zotero local API integration).
- Bilingual-first: Chinese→English expansion plus a built-in CN↔EN dictionary (300+ pairs).
- LLM-ready outputs: dual-format JSON items plus Markdown context_block for prompt consumption.
Findings snapshot
| Metric | Value | Note | Source |
|---|---|---|---|
| MCP tools available | 36 | Discover / Read / Write / Manage / Insight / Admin categories | PyPI |
| Text-cleaning rules | 52 | Removes boilerplate; reported avg 10.6% line removal | PyPI |
| Embedding backend (default) | ONNX INT8 (~347MB) | 2–3× faster on CPU vs FP32, <1% R@5 loss vs FP32 | PyPI |
| Minimum chunk floor | 200 chars | Prevents sub-fragment chunks that reduce retrieval accuracy | PyPI |
| Diversity tuning | MMR λ=0.4, max 3 chunks/paper | +54% paper diversity vs no diversification (reported) | PyPI |
| Evaluation suite | 60 golden queries | Recall@5/10/20, MRR, NDCG@10 benchmarking | PyPI |
How the Zotero RAG pipeline works (plain English)
zra-mcp ingests your Zotero library, extracts PDF text page-by-page (PyMuPDF), applies 52 regex cleaning rules, performs CJK-aware semantic chunking with IMRaD section detection, embeds chunks (bge-m3 quantized to ONNX INT8), indexes them in ChromaDB, and runs a hybrid search.
- Result format: JSON metadata plus a Markdown 'context_block' optimized for prompts (blockquote evidence plus relevance tiers).
- Bilingual expansion: dictionary plus Zotero tags, user synonyms, and OPUS-MT on-demand translation.
- Observability: JSONL retrieval traces, latency breakdowns, and index audit tools are available for reproducibility.
What this means for qualitative researchers
If you run literature reviews, coding-led syntheses, or multi-paper evidence mapping, zra-mcp buys you higher-precision discovery and LLM-friendly evidence clips, while leaving synthesis and cross-paper thematic quantification to your analysis pipeline.
- Faster evidence retrieval: switch from keyword hunting to meaning-based discovery, useful for concept-driven qualitative reviews.
- Better bilingual coverage: Chinese-English projects get integrated search without manual translation overhead.
- Reproducible traces: retrieval_trace and retrieval_stats support audits and protocol transparency for coding decisions.
Do more, faster with Evidano (map to the use case)
Problem: High-volume retrieval → Solution: Thematic synthesis
Import zra-mcp’s JSON items and Markdown context_blocks into Evidano to run thematic, frequency, and co-occurrence analyses across retrieved chunks and full papers.
Evidano automates hierarchical coding and shows code→subcode trees and co-occurrence networks so qualitative teams can move from dozens of hits to a validated theme map.
Problem: Multilingual evidence complicates coding → Solution: Transcription & translation controls
zra-mcp provides bilingual expansion and OPUS-MT translation; ingest both original chunks and translated text into Evidano to preserve nuance during analysis.
Evidano stores originals, translations, and provenance so teams can validate nuance and keep audit trails, and Evidano’s custom dictionary support helps keep domain terms consistent across CN↔EN during automated coding.
Problem: Inconsistent chunking and context → Solution: Expandable context + human-checked quotes
zra-mcp’s expand_context=True returns larger enclosing sections (~2000 chars) and Evidano preserves chunk metadata and surfaces clickable source quotes for rapid verification.
Evidano exports visual reports (word clouds, co-occurrence networks) that link back to the exact Markdown evidence blocks from zra-mcp for stakeholder-ready deliverables.
Problem: Sensitive or proprietary corpora → Solution: Secure-by-design workflows
zra-mcp runs locally; combine local retrieval with Evidano’s secure workspace when governance matters to keep sensitive data within controlled environments.
Evidano encrypts data in transit and at rest and does not use customer data to train third-party models, preserving governance and compliance for proprietary corpora.
Checklist: From zra-mcp retrieval to thematic insight (2-week pilot)
This checklist outlines the steps to run a compact 2-week pilot from Zotero search to a published memo.
- Week 0: Install pip install zra-mcp, enable Zotero local API, and run initial sync (note: first-run builds index; ONNX INT8 download ~347MB).
- Day 1–2: Run 20–30 representative queries (use the 60 golden queries as benchmarks) and save retrieval_traces for reproducibility.
- Day 3–5: Export top-100 context_blocks (JSON plus Markdown) and ingest into Evidano for automated coding and cross-segment analysis (by year, language, method).
- Day 6–10: Iterate codebook with Evidano’s AI-assisted coding; validate by sampling source blockquotes back in zra-mcp.
- Day 11–14: Produce stakeholder deliverables: heatmaps of theme frequency, co-occurrence networks, and an exec summary that links raw evidence back to Zotero records for transparency.
Wrapping up & next steps
zra-mcp is a strong retrieval and evidence-preparation layer for scholars who keep Zotero as their canonical library.
For qualitative teams who need thematic synthesis, cross-segment comparisons, and reproducible visual reports, pair zra-mcp’s local RAG outputs with Evidano’s secure ingestion, AI-assisted coding, and visual analytics to compress weeks of manual work into days.
Start by indexing a small Zotero collection with zra-mcp, export the top context_blocks, upload them to Evidano for a test analysis, and Try Evidano for free.
FAQ: AI-enabled Literature Review
What is zra-mcp and what does it do?
+
zra-mcp turns a Zotero library into a local RAG pipeline that extracts PDF text, cleans it with 52 regex rules, chunks with CJK-aware logic and IMRaD detection, embeds chunks with bge-m3 quantized to ONNX INT8, and indexes them for hybrid retrieval.
The pipeline outputs JSON metadata and Markdown context_blocks optimized for prompt consumption and evidence tracing.
Does zra-mcp run locally or in the cloud?
+
zra-mcp runs entirely on your machine using Zotero local API integration.
Local-first operation lets teams keep sensitive corpora on-premises while producing auditable retrieval traces.
How does zra-mcp support Chinese-English research?
+
zra-mcp supports bilingual-first workflows with a built-in CN↔EN dictionary (300+ pairs), bilingual expansion, Zotero tag and synonym support, and OPUS-MT on-demand translation.
This reduces manual translation overhead for projects that include Chinese and English literature.
How do I combine zra-mcp outputs with Evidano for thematic synthesis?
+
Import zra-mcp’s JSON items and Markdown context_blocks into Evidano to run automated coding, frequency analysis, and co-occurrence mapping across retrieved chunks and full papers.
Evidano preserves provenance and links back to source blockquotes so teams can validate themes and produce reproducible visual reports.
Topics
- AI-enabled literature review
- zra-mcp
- Zotero RAG
- ONNX INT8 embeddings
- bilingual literature search
Keep reading
- Commentary on NewsGhana's Obstetric Fistula Qualitative StudyLearn what Ghana's obstetric fistula qualitative study found from 10 interviews, five themes, and AI-enabled methods for safer health research. Explore Evidano.
- Commentary on NewsObstetric Fistula: Qualitative Analysis GuideLearn how qualitative analysis of obstetric fistula reveals barriers, stigma, and resilience in rural Ghana, and how AI helps health researchers synthesize interviews.
- Commentary on NewsWhat Qualitative Analysis Reveals About Obstetric FistulaLearn how qualitative analysis of obstetric fistula reveals stigma, care barriers, and resilience in Ghana, plus how AI can accelerate research synthesis.