Schema Crush

Tools 15

• schema_crush_biobert_match Match a biomedical term against candidate FHIR targets using BioBERT embeddings. Best for matching biomedical concepts and terminology where the source term carries clinical meaning. Returns calibrated confidence scores, not raw cosine similarity: scores are mapped through an isotonic calibrator fitted on 1,280 expert mapping rules, so 0.8 means roughly 80% empirical accuracy. Returns: JSON list of candidates with calibrated confidence scores, best first.
• schema_crush_explore_fhir_resource List the available fields on a FHIR R5 resource. Use to discover what a resource actually offers before mapping onto it, rather than guessing a field name that may not exist. Returns: JSON describing the resource and its fields.
• schema_crush_find_similar Find semantically similar mappings by vector search over the knowledge base. Use when lookup_mapping returns nothing and you want precedent: the nearest curated mappings serve as few-shot examples for an uncertain term. Returns: JSON list of similar mappings with similarity scores.
• schema_crush_get_feedback_stats Get statistics on collected human-in-the-loop feedback. Reports accept, reject, and correct counts so you can monitor whether the learning loop is receiving review signal. Returns: JSON summary of feedback counts and rates.
• schema_crush_get_transformation_rules Get FHIR transformation rules consolidated from GDC, CDA, HTAN, and ICGC. Explains the correct FHIR paths for common biomedical data patterns such as cancer staging, demographics, and specimen handling. Read these before proposing a mapping for anything non-obvious. Returns: The transformation rules for the requested section, as text.
• schema_crush_lookup_mapping Look up every curated mapping for a source term. Exact match, O(1). Searches the flat mapping database of expert-curated GDC and HTAN mappings and returns all matches rather than a ranked guess. Use this first: if the term is already curated, no inference is needed. Returns: JSON list of every curated mapping for the term.
• schema_crush_magneto_match Match a schema field name against candidates using Magneto structural embeddings. Best for matching field names by structure rather than clinical meaning. Trained on GDC to FHIR mappings. Returns calibrated confidence scores. Requires the optional magneto package, which is not on PyPI. If it is not installed this returns an error explaining how to install it; biobert_match and rule_match work without it. Returns: JSON list of candidates with calibrated confidence scores, best first.
• schema_crush_map_csv_columns Map a whole set of CSV column names to FHIR R5 paths in one call. Runs the full matching cascade per column and stops at the first stage that answers, which is what makes the result good: 1. knowledge base exact lookup, confidence 1.0 2. fuzzy match over curated field names 3. calibrated BioBERT embedding similarity 4. unmapped, flagged for human review Prefer this over calling biobert_match yourself. A single matcher used alone performs badly on field names: raw embedding scores for unrelated FHIR paths cluster within a few points of each other, so the wrong target often ranks first. The cascade avoids that by trying curated knowledge first, which resolves the large majority of real-world columns exactly. Uses no LLM and needs no API key. Returns: JSON list of {source, target, confidence, stage, note}, one per column,, plus a summary of coverage by stage.
• schema_crush_profile_csv Analyze CSV columns to infer use case, column groups, and mapping recommendations. Run this before mapping a new dataset. It interprets cryptic column names such as "death_event_1death_0censor" and reports the likely use case (survival analysis, variant tracking, and so on) plus per-column content types and FHIR recommendations. Returns: JSON profile with use case, column groups, content types, and, recommendations.
• schema_crush_record_feedback Record a human review decision on a proposed mapping. Feeds the learning loop that recalibrates matcher confidence. Writes are disabled on the public Space because the store is shared across all visitors; run your own instance with ENABLE_FEEDBACK_WRITES=1 to collect feedback. Returns: JSON confirmation, or an error if writes are disabled.
• schema_crush_rule_match Match a term against candidates using the curated expert knowledge base. Highest precision of the three matchers because it draws on expert-curated HTAN and GDC mappings rather than inference: 94.7% accuracy on field matching and 99.6% on content. Try this before the embedding matchers. Returns: JSON list of candidates with calibrated confidence scores, best first.
• schema_crush_search_fhir_fields Search for FHIR fields matching a keyword across all resources. Use when you know what a field means but not which resource holds it. Returns: JSON list of matching FHIR field paths.

All 15 tools