Skip to main content
Sphinx post-trains agents on your organization’s data and documents. It generates synthetic tasks grounded in those sources, evaluates agents attempting the tasks, and uses the results to refine the context that guides future agent behavior. The training output is external to the model’s weights: learned context that captures the definitions, relationships, procedures, and exceptions agents need to succeed. Your team can inspect and edit that context, and use it with different underlying models.

The post-training loop

  1. Provide training material. Ingest your data and documents, including data schemas, business rules, methods, slide decks, meeting notes, and other examples of your organization’s work. You can also train programmatically from Python using sphinx_learn.fit().
  2. Generate tasks and agent rollouts. Sphinx creates synthetic tasks from the training material and evaluates agents’ attempts to complete them.
  3. Learn from the results. Successful and unsuccessful attempts provide evidence about useful context, missing information, and incorrect assumptions. Sphinx uses that evidence to update its learned context, then repeats the process.
  4. Apply the training to new work. At inference time, agents use the learned context to interpret requests, find relevant evidence, and handle domain-specific exceptions.
  5. Evaluate generalization. Test on held-out tasks that were not used during training. Measure whether the agent applies what it learned to new questions, alongside cost and latency.
Try the process yourself with Train AI on a Business, using Sphinx, a runnable example using a fictional lender’s records and SQL data.

What the agent learns

The objective is better task performance. Training can teach an agent which business definition applies, how concepts relate, what evidence a decision requires, and when an apparent match falls outside a rule. For example, answering a question about customer value may require distinguishing a historical projection from a realized outcome, selecting the right measurement horizon, and recognizing exceptions for particular customer groups. The useful training output is context that helps the agent make those distinctions on a new task.

Inspectable and governable post-training

Because the learned context is external to model weights, you can see what Sphinx has learned, check it against your sources, and correct it directly. When your sources disagree, Sphinx helps agents recognize the uncertainty instead of giving a falsely confident answer. Use Benchmarks (beta) to check answers against your criteria and track regressions as the context evolves. Contact Sphinx to activate Benchmarks for your workspace.

Choose how your agent uses the trained context

We recommend the Responses API, especially when deploying an agent to your users. It integrates Sphinx’s learned context directly into the inference request path. MCP is easier to get started with for testing and internal use cases. Your agent chooses when to consult Sphinx through tool calls, which may add latency. Sphinx can also support private inference through a Private Gateway. Contact us to discuss your deployment.

Augment your existing data stack

Sphinx does not train on raw numerical data. It learns the context needed to use that data correctly: schemas, metric definitions, business rules, relationships between tables, and how to interpret results. Your agent uses its existing data connectors to query the underlying numbers.

Example: an agent connected to Snowflake

Suppose a user asks your agent, “What was our net revenue last quarter?”
  1. Sphinx provides the context. The agent uses Sphinx’s learned context to identify the right Snowflake tables and columns, understand your definition of net revenue, and apply the correct joins, exclusions, and fiscal-quarter boundaries.
  2. Your existing Snowflake connector runs the query. The agent constructs the query and executes it through the Snowflake connector already configured in your client or agent. Snowflake returns the numerical results; Sphinx does not replace that connector or supply the underlying numbers.
  3. Sphinx helps the agent interpret the output. The learned context helps the agent explain what the result means, account for business-specific caveats, and answer the user using the queried figures.

Work with source documents

The same approach applies to documents. For example, Sphinx can help an agent understand which contract terms matter and how your team interprets them. The agent uses its existing document connector or file tools to open the original contract, check the exact wording, and cite the evidence in its answer. Keep your agent’s database and document connections available for tasks that require live data or original evidence. Sphinx supplies learned context alongside those tools. If you provide files locally, tell the agent where to find them in its system instructions. Connecting Sphinx does not itself give the agent access to files on your computer or credentials for your databases.

Learning organizational context

Explore context-based post-training and generalization.

Inspect and govern learned context

Trace answers, review updates, and correct errors.