Skip to main content
Back

Parallel Web Systems Retunes Search APIs to Beat Rivals on Healthcare AI Accuracy

Parallel retuned its search index for FDA labels, trial registries and PubMed papers. In its own tests on four healthcare benchmarks, it was the most accurate search engine and often the cheapest per correct answer.

Repo, Full Access
  • Parallel for Healthcare upgrades search ranking for FDA labels, ClinicalTrials.gov, CMS registries and PubMed papers.
  • Parallel was the most accurate engine on four healthcare benchmarks, testing against Exa, Perplexity, Tavily and Claude's built-in search.
  • On PWS-Health: 9.8 correct answers per dollar, 1.4x the next-best engine and 6x cheaper than built-in search.
  • PWS-Health is Parallel's own private benchmark, so outsiders can't check it. On TrialQA, accuracy was nearly tied with Perplexity.
  • AcuityMD verified 574K records at about one-fifth of a frontier model's cost.
  • Works across all APIs at no extra charge; an open-source recipe turns a drug CSV into cited PDF profiles.

Healthcare agents can retrieve the correct paper and still produce a wrong answer when the needed evidence sits in a safety table, trial registry field, or methods section. Parallel Web Systems has launched Parallel for Healthcare, an update to the index and ranking behind its APIs. An index determines which sources are available for retrieval, while ranking determines which documents and passages the agent sees first. Parallel says the update improves access to FDA drug labels, ClinicalTrials.gov records, CMS registries, and recent papers indexed by PubMed.

The release introduces no separate plan, subscription, or integration path. According to Parallel, existing calls to its Search, Extract, and Task APIs automatically use the healthcare ranking and require no code changes.

The benchmark changes one variable

Parallel says it tested six healthcare benchmarks while holding the agent, questions, and grader constant. The setup was designed to isolate the effect of the search provider:

  • The answer model was Claude Opus 5.5 with the high-effort setting.
  • Each search engine ran three times on the same questions.
  • Total cost included answer-model tokens and search fees at list price.
  • Parallel audited public answer keys and applied corrected gold answers uniformly across providers.

The launch materials emphasize four benchmarks that cover different retrieval patterns:

  • PWS-Health: A private set of 133 held-out questions with verified answers, mostly adapted from customer workflows.
  • LitQA3: A component of FutureHouse’s LAB-Bench 2. Its 168 questions ask for details from recent biomedical papers, including effect sizes in results tables and steps in methods sections.
  • TrialQA: Questions about ClinicalTrials.gov records, such as a study’s control arm or planned enrollment.
  • MedBrowseComp: Multi-step questions about FDA drug approvals, with each retrieved answer informing the next search.

The most detailed cost comparison covers PWS-Health. Correct answers per dollar are calculated from accuracy and the reported cost per 1,000 questions.

Parallel’s reported PWS-Health results
Engine Accuracy Cost per 1,000 questions Correct answers per dollar
Parallel Advanced 83.7% $85.3 9.8
Parallel Fast 85.2% $104 8.2
Perplexity 80.5% $116.2 6.9
Exa 82.5% $124.6 6.6
Tavily 77.2% $154.3 5.0
Claude native search 66.9% $646 1.0

Parallel Advanced produced 9.8 correct answers per dollar, 42% more than Perplexity’s 6.9. Parallel Fast posted the table’s highest accuracy at 85.2%, although its higher cost reduced its efficiency score. Parallel Advanced’s listed cost was about 13% of Claude native search’s cost.

On LitQA3, both Parallel modes scored 86.7%, compared with Exa’s 81%, a lead of 5.7 percentage points. The result supports Parallel’s central claim that healthcare retrieval depends on locating the relevant passage inside a source, especially when the answer appears outside the abstract.

A paper match can still miss the evidence

Parallel’s LitQA3 examples show how passage selection changes an answer. One question asked how many patients in evorpacept’s first-in-human trial stopped treatment because of drug-related adverse events. Parallel found the 2021 Lancet Oncology paper in two searches and returned the correct answer: six patients. The competing engines answered one patient after drawing from a later lymphoma paper.

Side-by-side search comparison of Parallel, Exa, Perplexity, and Tavily on a clinical trial adverse-events question
Parallel’s published comparison of four search providers on a LitQA3 question.

In a second LitQA3 example, Exa found the correct paper twice, but the fetched passages omitted the methods section. The agent then answered, “I don’t know.” For agent builders, passage coverage within a document is a separate metric from document hit rate.

The evidence has clear limits

PWS-Health’s 133 questions and answers are private, which prevents outside researchers from reproducing the exact evaluation. Parallel also removed questions that every provider answered correctly, shaping the set toward harder and more discriminating items. The company acknowledges that vendor-built benchmarks can favor the vendor that designed them, and no independent analysts or publications had released their own tests at the time of writing.

The reported margins also vary by benchmark. On TrialQA, Parallel Fast scored 96.3% and Perplexity scored 95.2%, a difference of 1.1 percentage points. Cost produced a wider separation: Parallel reports that its result cost about one-ninth as much as Claude native search. Across the four highlighted tests, Claude native search was the least accurate and most expensive option in Parallel’s results.

These scores evaluate retrieval-supported question answering under one agent configuration. They provide no evidence about patient outcomes, diagnostic safety, or regulatory fitness.

A 574,000-record customer workload

Parallel’s production example comes from AcuityMD’s case study. AcuityMD operates a platform used by more than 500 medical-technology companies. Before adopting Parallel, its team found that frontier models cost about five times more and returned broken links for 5% to 10% of URLs.

Using Parallel, AcuityMD processed 574,000 records over five months at roughly one-fifth of the frontier-model cost. The workflow checked facts such as whether clinics remained open and which health system owned each office. Parallel published the case study, and the announcement cites no independent audit of its figures.

Diagram of healthcare agent workflows for drug discovery, provider enrichment, and provider screening using Parallel APIs
Parallel’s examples of healthcare workflows built with its APIs.

A constrained recipe for drug profiles

Parallel has published an open-source drug-profile recipe that demonstrates a source-restricted workflow:

  1. Run the research workflow through the Task API.
  2. Use Source Policy to restrict retrieval to regulator sites such as the FDA and EMA.
  3. Query ClinicalTrials.gov and PubMed through free Data Connectors.
  4. Validate trial and paper identifiers against their corresponding public records.
  5. Generate a cited PDF for each drug.

Parallel’s Data Connectors catalog also lists free access to bioRxiv, ChEMBL, CMS Coverage, medRxiv, the NPI Registry, and PubMed. These connectors give agents structured routes into domain-specific sources while Source Policy controls which sources a workflow may use.

Deployment claims to verify

Parallel advertises the following enterprise capabilities for regulated teams:

  • SOC 2 Type 2 controls
  • Zero data retention
  • A business associate agreement for HIPAA-covered workflows
  • EU data residency
  • Single sign-on
  • Specialized healthcare indexes

Plan availability, contractual scope, retention behavior, and regional coverage require confirmation during procurement. The healthcare release follows similar domain-specific upgrades for legal and financial search, extending Parallel’s strategy of tailoring retrieval to specialized source sets.

A fair test in your stack

An in-house evaluation can determine whether Parallel’s reported gains carry over to a specific clinical, pharmaceutical, or medical-technology workflow:

  1. Build a held-out set from representative production questions, including items whose answers appear in tables, methods sections, labels, and registry fields.
  2. Keep the answer model, prompt, grader, tool budget, and retry policy fixed across search providers.
  3. Run each configuration multiple times to capture variation.
  4. Measure answer accuracy, citation support, abstentions, broken URLs, latency, and total cost.
  5. Manually inspect high-risk answers and verify every cited passage against the primary source.

Keeping those variables fixed isolates retrieval quality and shows whether the healthcare ranking improves the questions that matter in production.