MOLECULAR ASSETS
Technology

Detailed ChemQuery Architecture

Our Architecture: Prompt, ChemQuery app, LLM and context files connect to a database of training data, supplemental data, and in-house content
Reading the diagram

A walkthrough, step by step

  1. Ask a question

    A scientist types a prompt into the ChemQuery app. Other applications, such as Scaffold-Lab, ChemAudit, Chemdoc miner, Claude-Science, ChatGPT and Gemini, can send prompts too, through MCP connectivity.

  2. The LLM plans the query

    The language model reads the prompt and writes an expert SQL query against the trained data structures. It works in both directions with the database: it consults the tables, then runs the query.

  3. Training tables ground the model

    About 70 tables are used for training, including PubChem, ChEMBL, Target Central, clinical trials, FDA, EPA, worldwide IP, SureChEMBL, DrugCentral, and Google data. New tables can be added.

  4. Supplemental tables extend the reach

    More than 200 supplemental tables are available, such as OTP, NCI (SAVI), Medline, ChEMBL 37, REACH, and ECHA. They are drawn in when a question needs them.

  5. In-house content stays yours

    In-house tables, paid data, and custom tables sit in their own area of the database. They can be added to training so the model learns their structure.

  6. The result comes back

    The query runs in the BigQuery database, and the LLM returns the result to the app as data, counts, and tables you can inspect.

Curated foundations

Normalized scientific datasets and cross-source relationships support questions that span disciplines.

Supplemental data

Specialized tables and additional sources extend the questions the platform can address.

Enterprise data

Customer-owned information may be incorporated under an agreed integration and access model.