Scientific Information Retrieval System™
Retrieval built for how scientists ask questions
SIRS reads abbreviations, asset codes and drug classes the way your team does, so the evidence it returns matches the question that was asked.
How SIRS works
How does SIRS work?
SIRS handles a scientific question in four steps, from interpreting what is being asked to passing only grounded insights to Causaly's agents. Standard retrieval pipelines rely on similarity search alone, while SIRS combines lexical, semantic, knowledge graph and multimodal retrieval and ranks the results before any answer is written, with each step calibrated by biomedical information specialists.
Understand the question
SIRS reformulates ambiguous or abbreviated queries, extracts key biomedical entities and recognizes intent. Curated ontologies then expand the query, so a class such as anti-CD20 monoclonal antibodies reaches all 28 drugs in that class.
Retrieve the evidence
Four retrieval methods run together, so evidence is found whether it matches by exact term, by meaning, by biological relationship, or sits inside a table or figure.
Fuse and rank
Results from every method are fused into a single list and ranked by factual accuracy and relevance to the biomedical context of the question, so the strongest evidence leads.
Synthesize a grounded answer
Insights are extracted from each document and kept only when the cited excerpt supports them, which gives the final answer sentence-level citations that link back to the source.
FXR activation represses CYP7A1 1 and reduces conversion of cholesterol to bile acids 2, a mechanism linked to raised LDL-C in FXR agonist trials 3.
Why does retrieval decide answer quality?
Common questions have abundant, easily matched evidence. The questions that shape R&D decisions sit in the long tail, combining mechanisms, populations and sources that rarely appear together, and an agent can only reason over the evidence that retrieval returns.
Like-for-like benchmark with identical documents and twenty scientific questions (51 expected evidence units), comparing Causaly with an AWS OpenSearch baseline configured to industry best practice.
How does SIRS keep unsupported claims out of the answer?
Every insight is checked against the excerpt it came from before synthesis, and each statement in the final answer links to the highlighted passage in its source, so reviewers can verify it in seconds.
Extract insights
Relevant insights, not raw chunks, are taken from each document.
Ground and validate
Each insight is mapped to a cited excerpt, and unsupported ones are removed.
Classify the evidence
Direct answer, related evidence or no answer, which keeps limits visible.
Synthesize with citations
Supported insights form one answer with sentence-level citations.
Clinical program evaluation
“If it outputs a table, I can pull the exact references behind each point. It’s faster and reduces the risk of missing key perspectives.”
What evidence does SIRS search?
SIRS searches the public biomedical record and your own scientific documents with the same pipeline, and new publications become searchable within 24–48 hours.
The same retrieval over your internal documents
Investigator brochures, study reports and toxicology reports are parsed down to the section, table, row and figure in a dedicated single-tenant environment.
Frequently asked questions about SIRS
What is the Scientific Information Retrieval System (SIRS)?
SIRS is the retrieval layer of the Causaly platform. It interprets a scientific question, retrieves evidence from public and internal sources through lexical, semantic, knowledge graph and multimodal retrieval, ranks that evidence, and passes only grounded insights to Causaly’s agents.
How is SIRS different from standard RAG or vector search?
Standard RAG pipelines rely on vector similarity alone and return a short list of loosely related passages. SIRS adds exact-term matching for asset codes and study IDs, ontology expansion for synonyms and drug classes, Bio Graph retrieval for mechanisms, and table and figure retrieval, then ranks the combined results before any answer is written.
How does SIRS reduce hallucinations?
Each insight is mapped to a cited excerpt from its source, and claims without support are dropped before synthesis. Evidence is also classified as a direct answer, related evidence or no answer, so the response states its own limits.
Can SIRS search our internal documents?
Yes. Private Data deployments run SIRS over internal documents such as investigator brochures, clinical study reports and toxicology reports in a dedicated single-tenant environment, retrieving down to the section, table, row and figure with provenance preserved.
How current is the evidence SIRS searches?
New publications are searchable within 24 to 48 hours of publication and are added to the Bio Graph within about four weeks.
Does Causaly train models on customer data?
No. Customer data is not used to train or fine-tune models, and model providers operate under zero-data-retention terms.
Go deeper into how SIRS retrieves and grounds scientific evidence
The whitepaper covers the full retrieval architecture and benchmark method, and Causaly lets you see SIRS-powered answers on your own research questions.
