Finding the right circuit, whether it’s drawn, described or coded

Why search across formats?

Analog-circuit designers often reuse known designs, but the same circuit may appear in a repository as a schematic image, a written explanation or a SPICE netlist: a text-based description of components and how they are connected. A search system that understands only one format can miss useful examples in the others. Searching by keywords is also difficult when someone does not know the right terminology.

Yihan Wang and colleagues introduce AnalogRetriever, a system intended to search between all three formats. A user could, for example, enter a description and retrieve a schematic or netlist, or submit a diagram and search for related descriptions. The researchers also tested whether retrieved designs could help an AI system generate functional circuits.

Repairing the training material

To teach the system that different representations can refer to related circuits, the researchers assembled aligned examples from the Masala-CHAI dataset. They report that many original netlists had errors: only 22% compiled with the Ngspice simulator, and 11.4% passed a check of their direct-current operating point.

The team used a two-stage repair process. A language model proposed fixes, and Ngspice checked whether the resulting netlists compiled and passed the operating-point test. The researchers also used the process to fill in missing netlists or descriptions. The resulting collection contained 6,371 triplets linking a netlist, schematic and description. The reported compile rate was 99.97%, while 99.7% passed the operating-point check. These checks improve the consistency of the training examples, but do not by themselves establish that every circuit is useful for every design purpose.

Three formats, one search space

AnalogRetriever uses different tools for different inputs. It encodes schematic images and written descriptions with a vision-language model, and represents netlists as graphs. In that graph, devices and electrical connections are nodes, while the type of device terminal connecting them is recorded in the edges. This matters because, for example, a transistor’s gate and drain have different roles.

The system learns to place related examples from all three formats near one another in a shared mathematical space. Its training also gradually introduces harder comparisons between circuits with similar functions but different structures, aiming to help it distinguish designs that might otherwise look alike in a broad description.

The researchers evaluated six directions of cross-format search—for example, description-to-netlist and netlist-to-schematic—on held-out examples. They report an average Recall@1 of 75.2%. Recall@1 means that the top-ranked result matched the test pair under their evaluation setup; it does not mean that 75.2% of all real-world searches will find a suitable design. The test used galleries of 200 items sampled from 1,000 held-out examples, and the paper notes that several circuits might be valid matches for a single query.

For comparison, the strongest reported zero-shot baseline averaged 4.7% Recall@1 on the same evaluation. This is a large difference in this particular test, not evidence that the system will outperform every search tool or work equally well on other circuit collections.

Does retrieval help circuit generation?

The team also added retrieval to AnalogCoder, an AI-based circuit-design framework. Across eight language models and a 24-task benchmark, each model had a higher reported functional pass rate with retrieval; the average rose from 62.0% to 67.6%. A functional pass means meeting the benchmark’s checks, not proving that a design is ready for manufacture or suitable for every application.

This experiment suggests that giving a generator relevant circuit examples can help it produce working designs on the tested tasks. It does not isolate every possible cause of the improvement: the pipeline also expanded short task prompts into more detailed descriptions before retrieval. The benchmark is small, and the paper itself identifies limited coverage—19 canonical analog topologies—as a constraint. Mixed-signal, radio-frequency and power-management circuits are among the areas it says remain to be explored.

What to take away

AnalogRetriever offers a practical approach to a real engineering problem: connecting circuit information that is stored in different formats. Its reported results support further testing, particularly on broader collections and less familiar designs. They do not establish performance on proprietary industrial repositories, or show that retrieval can replace engineering review and simulation.

There is also a publication-status caveat. The supplied source is an arXiv preprint, while the extracted paper text includes conference-proceedings details for ICCAD ’26, dated November 2026. Those details cannot be reconciled from the available source record, so readers should check the linked arXiv page and publication record for the paper’s current status. Preprints may not have undergone peer review, and results can change as work is reviewed or reproduced.