Skip to content

ADVANCED_RESEARCH_WORKFLOWS.md

Latest commit

 

History

History
107 lines (82 loc) · 5.06 KB

ADVANCED_RESEARCH_WORKFLOWS.md

File metadata and controls

107 lines (82 loc) · 5.06 KB

Advanced Research Workflows

This page makes the newer research surfaces visible from the docs-site navigation: research timeline/lineage tree, Open-i image search, uploaded-image handoff, and persistent query memory.

Quick Map

Need Start here Continue with
See how a topic evolved over time build_research_timeline analyze_timeline_milestones, compare_timelines
Find biomedical images from text search_biomedical_images get_article_figures, unified_search
Upload or pass an image and search by its meaning analyze_figure_for_search search_biomedical_images, unified_search
Re-open large search/fulltext outputs without rerunning read_session(action="artifact") read_session(action="list_artifacts")

Research Timeline / Lineage Tree

Use timeline tools when the user asks how a field changed, which papers look like milestones, or how two research tracks compare.

build_research_timeline(topic="remimazolam ICU sedation", output_format="tree", max_events=20)
build_research_timeline(pmids="12345678,23456789", topic="Selected studies", output_format="mermaid")
analyze_timeline_milestones(topic="CAR-T therapy")
compare_timelines(topics="remimazolam,propofol,dexmedetomidine")

build_research_timeline accepts either a topic or a known PMID set, including explicit comma-separated PMIDs. It supports text, tree, mermaid, mindmap, json, json_tree, timeline_js, and d3 output formats. For a lightweight preview inside a normal search response, use unified_search(options="context_graph"); it is a preview from the current PMID-backed ranked set, not a complete graph. The planned persistent/versioned Research Chronicle is specified in Research Chronicle Rebuild Spec.

Open-i Biomedical Image Search

Use search_biomedical_images when the visual finding is already text and the goal is image evidence from Open-i.

search_biomedical_images("chest X-ray pneumonia", sources="openi", image_type="x", limit=10)
search_biomedical_images("histology liver fibrosis", sources="openi", image_type="mc", license_type="by")

Open-i expects English medical terminology. For non-English prompts, the agent should first translate anatomy, finding, and modality into English, then call search_biomedical_images. Open-i is strongest for radiology, microscopy, clinical photos, and teaching images; for article-native figures, use get_article_figures on PMC Open Access articles.

Uploaded Image To Literature Search

analyze_figure_for_search is the handoff tool for images supplied by an MCP client. It accepts an image URL or a base64/data-URI image and returns MCP ImageContent plus instructions for the LLM agent.

analyze_figure_for_search(image="data:image/png;base64,...", search_type="medical")

The server does not perform standalone visual diagnosis. The intended workflow is:

  1. The MCP client passes the uploaded image or image URL to analyze_figure_for_search.
  2. The LLM agent uses its vision capability to describe the image and extract English biomedical search terms.
  3. The agent immediately continues with search_biomedical_images for similar biomedical images or unified_search for related papers.

Persistent Query Memory

When session persistence is configured, large unified_search and get_fulltext outputs can be saved as artifacts. The immediate tool response can stay compact while the reusable payload remains available for later reads.

read_session(action="list_artifacts")
read_session(action="artifact", artifact_id="...")
read_session(action="artifact", artifact_uri="artifact://...")
read_session(action="artifact", artifact_uri="artifact://...", artifact_file="audit.json")
read_session(action="artifact", artifact_uri="artifact://...", artifact_file="results.json", offset=0, max_chars=200000)

Artifacts are query memory, not a second search. Reading them does not rerun external source calls. Local filesystem paths are redacted by default because remote clients cannot read the server host path. Set PUBMED_ARTIFACT_INCLUDE_LOCAL_PATHS=true only for local MCP clients that really need local_path and manifest_path.

For unified_search, artifact files are intentionally richer than the immediate MCP response. Read audit.json first for completeness warnings, then query_strategy.json for the executed source/query plan, then results.json or results.toon for the full article list. This keeps response tokens small while leaving enough evidence for repeated agent reads, sandboxed clients, and future remote artifact backends.

Verification Status

The current primary 46-tool MCP server exposes these tools directly:

  • Timeline: build_research_timeline, analyze_timeline_milestones, compare_timelines
  • Image search: search_biomedical_images
  • Uploaded-image handoff: analyze_figure_for_search
  • Query memory: read_session(action="artifact")

Coverage is guarded by docs alignment tests, tool registry tests, image-search tests, vision-search tests, timeline tests, and session artifact tests.