Introducing Node 1.1
Point-of-care search built on new medical search models, with a stronger Indian library and a new evidence filter.
The Cynaps team, NBALL Technologies India Private Limited2 min read
Today, we're introducing Node 1.1, the next version of the model behind Cynaps' point-of-care answers. Node 1.1 searches with new medical search models, keeps Indian guidance from being crowded out as the library grows, and picks the evidence for each answer in a single step. Node 1.1 is in testing and not yet available to physicians.
Search built on new medical search models
Node 1.1 searches with new medical search models in place of Node 1.0's. Across 315 development questions, Node 1.1 put every passage needed for the answer in the top 90 for 273 questions, against 268 for Node 1.0. That difference is within chance, so we describe it as "not worse", not "better".
A larger Indian library that holds its ground
As we added national treatment guidelines from MoHFW, they began pushing older ICMR and AIIMS passages out of the results. The obvious fix hid a bigger loss: reach on MoHFW questions fell from 29 of 30 to 16 of 30. Node 1.1 removes duplicate passages from the same document and gives new Indian sources a lane of their own. On the question sets we used to choose these fixes, it reached 277 of 315 overall, 74 of 81 Indian-source questions and 27 of 30 MoHFW questions. We now report reach for each source, not only overall.
An evidence filter that keeps what answers the question
Before writing, Node chooses which passages the writer reads. Node 1.1 reviews all 90 candidates together and keeps the passages that state the answer, so the writer works only from evidence that answers the question. It kept the same needed passages as before (263 vs 265, within chance).
Evaluation
An internal team that did not build Node 1.1 wrote and sealed a fresh exam. Node 1.1 got the needed evidence to the writer for 55 of 65 scored items (84.6%; 95% interval 73.9–91.4%).
Limitations
In 8 of the 10 missed items, the writer was given a passage with the right words that belonged to a different product, formulation or patient group. That's where most of our next work goes. These results measure whether the right evidence reaches the writer, not how accurate every answer is.
Availability
Node 1.1 is in testing and not yet available to physicians. We'll announce here if and when that changes.
What's next
We plan to have physicians review our test questions, and to publish a benchmark so others can check our numbers. Future fine-tuning will use synthetic questions written from our library and reviewed by physicians. It will not use patient data or the content physicians enter in Cynaps.