Evidence-based guideline retrieval

The useful output is a retrieved passage that names the guideline document and the section it came from. If the indexed corpus has nothing on this question, the field stays empty. Empty is a result. A fluent paragraph with no cite is not.

Hospitalists already know how to read a guideline. The problem at assessment is location under time pressure: which institutional pathway applies to this presentation, which society document is current on your service, which section actually addresses the decision in front of you. Retrieval should solve that location problem. It should not write the recommendation.

Treat a result as complete only when you can hold all of this at once: the passage text, the source title, a version or effective date if the index stores one, and a section heading or numbered item you can open in the original. Reject a synthesized "consider X" with no document. Reject a blend of two pathways presented as one protocol. Reject a fill-in that sounds like your shop's sepsis bundle because similar language exists in a different hospital's PDF.

You still decide. The passage is evidence you can check against the chart. It is not an order, a care plan, or a substitute for the history, exam, and labs you already owe the patient.

Retrieval quality is decided before anyone types a query. Index the documents your medical staff, quality committee, or equivalent has named as in force: society guidelines you have adopted, locally owned bundles, and the care pathways that actually run on your floors. If a PDF is sitting on a shared drive and nobody owns the version, it does not belong in the index.

External knowledge follows the same rule. UpToDate-class sources can sit next to internal pathways when your license and governance allow it, and when the retrieval layer can still return a document identity and a section rather than a paraphrased monograph. Stores adjacent to Epic or Oracle Health, and enterprise search stacks in the Microsoft class, are plumbing. They do not turn an unvetted file into policy.

Versioning belongs in the index, not in a hallway comment. A 2019 heart-failure pathway and a 2024 revision must not collide silently. Tag each object with title, issuing body, effective date, and the section structure you expect to cite. When a document is retired, remove it or mark it historical so it cannot surface as current.

Scope the index to assessment questions you actually ask on the service: admission criteria, initial workup, first-line versus alternative therapy when the chart already constrains the choice, isolation and disposition rules, and when to escalate. Do not index every educational slide deck because a larger corpus feels safer. Noise at index time becomes confident wrong answers on a night admission.

Someone has to own additions and withdrawals. A guideline that is "in the system" with no steward is how last year's vancomycin timing language keeps ranking next to this year's pneumonia pathway.

Run retrieval while the assessment is still open: after you have a working problem list, not as a decorative cite at sign-out. Query the clinical question. Do not dump the entire chart into the prompt. Pair a short problem statement with the constraints you already know (pregnancy, dialysis, recent bleed, local antibiogram policy if that document is in the index).

Return format is non-negotiable. Each hit must carry passage text, document title, section or an equivalent locator, and enough version signal to know you are not reading a withdrawn item. Rank is not a cite. A similarity score is not a section number.

If nothing meets that bar, leave the field blank. Do not lower the threshold until a sentence appears. Do not ask the model to guess the pathway you usually use. Silence tells you to open the source yourself, call the consultant, or proceed on your own reading of the chart. That is slower than a fake snippet and safer.

This step sits next to other assess-stage work without replacing it. Use symptom-to-differential generation when you need a working list of what to consider. Guideline retrieval should only speak to documents you indexed. Use prior record ingestion and summarization for what already happened to this patient. It does not tell you what the current heart-failure pathway says.

A hospitalist is admitting a patient with community-acquired pneumonia, chronic kidney disease, and a documented penicillin allergy. The question at assessment is not what a general model would recommend. It is whether your indexed CAP guidance and your institutional antibiotic pathway say anything specific about beta-lactam alternatives in renal impairment, and which section to open.

A complete retrieval looks like this: a short passage from the named institutional pneumonia pathway, section on initial inpatient antibiotics, plus a second hit from the adopted society CAP guideline, section on allergy and alternative agents. Each hit includes document title and section heading. The hospitalist reads both, checks the allergy details and eGFR in the chart, and writes the plan. Nothing in the snippet is signed as an order.

An incomplete retrieval looks like this: a paragraph that says consider a respiratory fluoroquinolone with no document and no section. Or a blended paragraph that mixes your sepsis bundle's antibiotic timing language with pneumonia drug choice as if they were one pathway. Treat both as empty. Open the pathway yourself or proceed without a machine cite.

This is an illustration of the contract, not a measured outcome. Either the index contained those sections and the retriever returned them with locators, or it did not.

A recommendation with no section. Fluent advice that cites guidelines without a title you can search and a heading you can open is not retrieval. It is generation. Reject it the same way you reject an unsigned verbal order: go to the source or do without.

Treating the snippet as an order. The passage is not a medication-administration action, not a bundled order set, and not a substitute for allergy, renal, and interaction checks you already owe the patient. If the EHR can jump from a cite into an order set, that jump is a separate, governed workflow. Retrieval itself should stop at the cited passage.

Inventing a pathway. When the corpus is silent, a model can still write something that sounds like your shop: the right unit names, the right severity scores, a familiar antibiotic. That text is not a care pathway. It was not indexed, not approved, and not versioned. Leaving the field empty is the correct behavior. Filling it with a plausible local-sounding protocol is the failure.

A quieter related failure: retrieving the right document but the wrong section, or an outdated revision that still ranks well. The cite is how you catch that. If you cannot land on the section in one open, you do not have a cite.

You own the assessment, the plan, and the decision to treat, wait, or call. Retrieval can shorten the hunt for the document your institution already said you should use. It cannot absorb indication, contraindication, or the mismatch between a general section and this patient.

When the next question is drug choice under a known genotype, that is a different assess-stage task: pharmacogenomics-based drug selection, which still needs a cited source and still leaves the order to you. When the question is whether this patient meets criteria for an open study, that is clinical trial matching, not a substitute for the pathway you just retrieved.

If the index is thin, fix the index. Do not compensate with a more talkative model. If the index is rich and retrieval still returns uncited prose, fix the return contract: passage, document, section, or blank. The quality outcome is that contract, held at the bedside, every time.