Skip to main content
Docs

Search guides and API endpoints, for example “Idempotency-Key” or “submit job”.

    Concepts

    The Assistant

    How the Assistant plans and prices runs, why it can only propose them, which quick calculations it runs itself, and what each turn costs.

    Updated October 8, 2026

    On this page

    The CogniChem Assistant plans computational work from a plain-language request. It runs in the app; Use the Assistant is the quick start.

    Confirm to run

    The Assistant can read, look things up, check, and estimate, but it cannot start a job or workflow, or cancel or delete anything. When it has a plan it proposes a run, which appears as a plan card with the server's estimate. Only your Approve starts it:

    1. Approving checks the plan again and re-prices it at your current rates.
    2. If the price went up, nothing starts; the card shows the new price for you to approve.
    3. If your wallet can't cover it, nothing starts and the card stays pending.
    4. Otherwise the job or workflow starts exactly as if you had submitted it yourself, under your account.

    Each card can be approved once, and expires after 30 minutes. Approving costs no tokens.

    Quick calculations

    For a quick question about a few structures, the Assistant answers in the reply instead of proposing a run. It can convert formats and get identifiers (InChI, InChIKey), standardize structures, check Lipinski and Veber rules and PAINS / Brenk alerts, compute the RMSD between two structures, look up fluid properties at one state, and predict a property with a public MPNN model or one you trained.

    • These are not charged to your wallet. Each utility call counts once, and each predicted molecule once, against your plan's monthly allowance, the same as on the tool pages.
    • Each message can run up to 4 calculations and 1 prediction (up to 100 molecules); each conversation up to 40. A calculation that takes longer than 30 seconds is cancelled and not counted.
    • Structures must come from you, from a lookup, or from an earlier calculation, and are screened for controlled and chemical-weapons substances first, like any run. Coordinate files are read from your library by reference.
    • The reply labels values as computed by CogniChem or predicted by a named model, with that model's validation metrics. They are not database or experimental values.

    What it can look up

    • CogniChem's catalog: tools, job schemas, workflow templates, and ports.
    • Public databases: PubChem (compound names to structures), ChEMBL (molecules, targets, activities), Open Targets, UniProt, and the RCSB PDB, through CogniChem's lookup service. Results carry a link to their source record and licence.
    • Method papers for each tool.
    • Your results: the status of runs started from the conversation, and records from their output files.

    Only names and IDs are sent to the public databases. ChEMBL similarity and substructure search need a structure (SMILES) to leave CogniChem, so they work only after you turn on structure search for that conversation. Files you attach stay in your artifact library: the Assistant passes them to tools by reference and never reads their contents into the conversation.

    What it costs

    Turns are charged to your wallet for the language-model tokens they use, at your plan's rates:

    1. Before a turn, CogniChem holds enough for the longest reply it could give (a turn can make several model calls, for example to use tools).
    2. After the turn you pay for the tokens actually used, and the rest of the hold is released.
    3. Stopping a turn, or closing the page during one, charges only for what was used so far. Turns that fail on the provider's side are not charged.
    CogniChem Assistant, Low reasoning: USD per million tokens
    TokensBasicStarterProEnterprise
    Input$0.70$0.50$0.40$0.30
    Cached input$0.070$0.050$0.040$0.030
    Output$4.20$3.00$2.40$1.80
    CogniChem Assistant, High reasoning: USD per million tokens
    TokensBasicStarterProEnterprise
    Input$7.00$5.00$4.00$3.00
    Cached input$0.700$0.500$0.400$0.300
    Output$35.00$25.00$20.00$15.00

    Limits

    • Spend cap: each conversation can spend up to $1.00. When the next turn would pass it, the Assistant stops and asks; continuing raises that conversation's cap by $1.00.
    • Quick calculations: 4 per message, 1 prediction per message, 40 per conversation (see above).
    • Time: a turn has 180 seconds in total, tool calls included. A turn cut off by the limit is charged for what it used and keeps its partial reply.
    • Message length: up to 16,000 characters.
    • Memory: each turn sends the model the most recent part of the conversation (about 12,000 tokens); older messages stay on screen but drop out of its context.

    From code

    The Assistant's API is under Assistant. It needs a signed-in access token (API keys are refused) and streams each turn as server-sent events. The Python SDK wraps it in client.chat.