Multi-step
Reasoning agent with tool selection, retrieval, integrity checks, and synthesis.
Tool-driven
Custom ES|QL tools, semantic search, and scheduled workflows rather than prompt-only output.
Domain-specific
Focused on scientific evidence verification where provenance and study quality matter.
Why this exists
Scientific claims are expensive to verify properly.
A reader sees a headline claiming that a diet intervention reverses ageing or reduces inflammation. Verifying that statement usually means searching papers, screening abstracts, weighing methodology, checking for retractions, and spotting funding conflicts.
SciVerify compresses that work into a repeatable agent workflow without hiding the reasoning steps. The result is not just an answer. It is an auditable chain of evidence.
99% faster review loop
Literature search and first-pass synthesis drop from hours to seconds for scoped claims.
Zero invented citations
Evidence comes from Elasticsearch retrieval and ES|QL outputs, not fabricated references.
Second-agent challenge step
BiasDetector checks for cherry-picking, funding signals, and quality gaps before the final verdict.
Demo
Recorded proof of the agent in action.
Claim verification workflow
Shows decomposition, semantic retrieval, ES|QL filtering, methodology assessment, integrity checks, and verdict generation.
Download MP4
Evidence quality analysis
Highlights quality-focused tools such as
find_high_quality_evidence and methodology distribution analysis.
Workflow
Five steps from claim to verdict.
Claim decomposition
Break the claim into subject, intervention, outcome, and population.
Evidence retrieval
Run semantic search and ES|QL filters in parallel to pull relevant literature.
Evidence evaluation
Score relevance, study type, citation strength, and support versus contradiction.
Integrity review
Flag retractions, weak sample sizes, funding signals, and outlier patterns.
Synthesis
Produce a verdict, confidence level, evidence map, and reasoning trace.
Architecture
Built around Elasticsearch as the retrieval and tool substrate.
The system uses Elasticsearch for indexed literature, semantic retrieval, and ES|QL analytics. Agent Builder orchestrates a primary verification agent and a peer-review agent that challenges weak conclusions.
- Elastic Agent Builder for agent registration and execution
- Custom ES|QL tools for filtering, analytics, and integrity checks
- Semantic Scholar ingestion pipeline for building the corpus
- Elastic Workflows for scheduled freshness and re-ingestion
Build notes
What this project demonstrates.
Hackathon fit
This project fits the Elastic Agent Builder brief closely: a context-driven, tool-using, multi-step agent with measurable operational value in a narrow domain.
Features used
Agent Builder agents, custom ES|QL tools, semantic retrieval, multi-agent review, and workflow-based ingestion freshness.
Challenges
Balancing retrieval precision with broad literature coverage, making reasoning visible, and preventing overconfident synthesis from sparse or biased evidence.
Explore the project