Retrieve authorized context before generating
Search your document store using the current user’s permissions. Select the passages needed for the question and pass them to the model as context with stable source identifiers.
Keep authorization in application code. The model should only see material the user is allowed to access, and a model response must not be used to decide that access.
Call the selected DeepSeek model
Configure the TextCortex base URL and choose a DeepSeek identifier available to your account. For European processing requirements, use the intended EU-hosted deployment.
Send the question, retrieved passages and instructions in your chat request. Tell the model to distinguish supported answers from missing information, then validate the output before presenting it.
Make citations verifiable
Ask the model to reference the source identifiers supplied with the passages. Resolve those identifiers to links in application code, and reject references that were not in the retrieved set.
This approach avoids letting generated text invent a trusted destination. It also lets you inspect whether a correct-looking answer is actually supported by the supplied material.
Compare model choices with the same retrieval layer
Keep the retrieved passages fixed while evaluating DeepSeek against GPT, Claude, Gemini, Kimi or GLM through TextCortex. Score factual support, refusal when context is insufficient, output validity and latency.
The DeepSeek EU API page covers model access. For the client setup, follow the OpenAI-compatible API guide.