DeepText¶

DeepText is a powerful prompt experimentation tool within DeepExtension that allows users to design, test, and compare structured prompts across different model states and sources. It is designed for business logic alignment, prompt evaluation, and reproducible model behavior — all without writing code.

What Makes DeepText Unique¶

Unlike traditional prompt playgrounds, DeepText provides tight integration with your models and embedded enterprise documents, enabling deeper control and more realistic testing environments.

Key features include:

Inference with Customized Models: You can run prompts directly against trained adapters or PEFT checkpoints — even before they are merged into a base model. During model selection, customized models and all third-party models are listed, and each is clearly labeled by type.
Embedded Document Integration: Reference enterprise documents or embedded knowledge bases directly inside your prompts. This is especially useful for retrieval-augmented generation (RAG) use cases.
System Prompt Control: You are strongly encouraged to define a System Prompt — this guides model behavior and improves inference quality. If left blank, a default system message like "I am an AI assistant." will be automatically injected.
Side-by-Side Model Comparison: Easily compare inference results from two different models by clicking “Add a model”. All prompt settings from the upper model can be cloned to the lower section using “Copy parameters”, enabling fair comparisons with minimal effort.
Realistic Production Simulation: Test partial or intermediate training outputs in a simulated production flow. This allows for early evaluation of prompt effectiveness before full model deployment or merging.
Multimodal Image-to-Text: Supports visual understanding tasks by generating textual descriptions from images, enabling applications like automated captioning and visual content analysis.

Important Notes¶

For customized models, the corresponding complete models (after merging) are expected to produce identical outputs during inference.
We strongly recommend using live models for production scenarios, as they offer significantly better inference speed and stability.
The inference capability for customized models is primarily intended for validation and experimentation, not for high-load or low-latency production usage.
Base models are not available for direct inference here. If you wish to use them, we recommend deploying them first to benefit from much faster performance.

Use Cases¶

Evaluate prompt robustness across model versions
Compare inference behavior across different configurations for testing across multiple dimensions.
Test business rules using real enterprise data
Validate document-aware inference
Tune prompts for internal tools or APIs

Next Steps¶

Select the model type
Start by selecting a model and input prompt
Embed reference data or system instructions as needed
Use the comparison mode to evaluate inference side by side
Share and log prompt tests with your team

DeepExtension — Precision prompting meets enterprise-grade model management