← Retour aux guides

Hallucination-Resistant Code Generation in Production

Why code models hallucinate in the first place

Code generation hallucinations usually happen when a model is asked to fill gaps it cannot actually see. It may invent a function name, guess the wrong framework version, or create a file path that does not exist. In production, that is expensive because one wrong suggestion can break a build, waste review time, or create subtle bugs that pass a quick glance.

The good news is that hallucinations are not random. They tend to show up when context is thin, instructions are vague, output format is loose, or the model is allowed to improvise too freely. If you treat code generation like a controlled workflow instead of a creative writing task, you can reduce errors dramatically.

Step 1: Give the model only the context it needs

Start with the actual source of truth: the relevant files, dependency list, failing test output, and any API contract the code must obey. Do not ask the model to infer your architecture from a one-line prompt. For example, if you want a refactor in a TypeScript service, include the service file, the interface it implements, and the test that fails. That way the model is less likely to invent a library call or guess the wrong type shape.

• Include exact file contents for the files being edited.
• Paste the error message, stack trace, or failing assertion.
• Provide the package manifest or lockfile snippet when dependencies matter.
• Tell the model to ask for more context instead of guessing.

Step 2: Use a strict prompt contract

Do not say, fix this bug and hope for the best. Instead, define the task boundary and the output shape. A strong prompt contract might say: edit only the listed files, do not create new dependencies, preserve existing public APIs, and explain any assumption explicitly. If the model cannot complete the task with confidence, it should say what is missing.

This single change prevents a lot of hallucinations. The model stops acting like a general advisor and starts behaving like a constrained code assistant. In practice, that means fewer fake imports, fewer imaginary helper functions, and fewer changes outside the requested scope.

Step 3: Ground generation in retrieved code

For anything beyond a trivial snippet, retrieve the real code before generating. A lightweight retrieval step can pull the relevant class, its callers, related tests, and nearby utility functions. This is especially important in larger repositories where naming collisions are common and a model may confidently select the wrong abstraction.

• Chunk files by function or class, not by arbitrary token count.
• Retrieve call sites and tests, not just the target file.
• Pass framework-specific conventions, such as routing rules or ORM patterns.
• Ask the model to cite the exact file names or symbols it used.

That grounding makes the response verifiable. If the model says it updated a method, you can confirm the method exists in the retrieved context before you accept the patch.

Step 4: Reduce randomness and add a reviewer loop

For code generation, low temperature usually helps. Start with a deterministic setting and only increase it when you need brainstorming rather than implementation. Then add a second pass that checks for unsupported APIs, incorrect file paths, and mismatched types. The reviewer can be the same model family or a different one, as long as it is asked to verify against the supplied context.

A practical pattern is: generate a patch, run lint and tests, then ask the model to review only the diff and the failing output. This catches hallucinations that look plausible in isolation but do not survive execution.

Step 5: Turn past hallucinations into test cases

The best evaluation set is your own failure history. Save examples where the model invented an import, used a nonexistent method, or edited the wrong file. Then build regression tests around those cases. Your eval does not need to be complicated. It just needs to answer one question: did the assistant stay inside the boundaries and produce runnable code?

• Check for nonexistent imports or symbols.
• Verify no new files are created unless requested.
• Run unit tests, type checks, and linting in CI.
• Score outputs for factual consistency with the source files.

Step 6: Make iteration affordable enough to do it right

Reducing hallucinations takes experimentation. You will compare prompts, try different models, and run repeated evals until the failure rate drops. That is exactly where 59API is useful. With one base URL, https://api.59api.com, you can plug into Claude Code, Codex, or any OpenAI SDK you already use, without changing your application architecture. It is pay-as-you-go, among the cheapest relays, and it gives you access to native official-quality Claude and GPT models with no downgrade, so your tests reflect real model behavior. If you want to burn fewer tokens while you tune prompts and build evals, this is a very practical place to start.

A production checklist you can use today

• Provide real repo context, not just a vague task.
• Require the model to say when it is uncertain.
• Retrieve relevant files before generation.
• Use low temperature for implementation work.
• Validate every output with tests and static analysis.
• Track hallucinations as failures and add them to your eval set.

When you combine these steps, hallucinations stop being mysterious and become measurable. You will still get the occasional miss, but they will be easier to detect, cheaper to correct, and far less likely to reach production. If you are building an AI coding workflow and want to test it at a realistic cost, sign up for 59API and run your next prompt-eval cycle on real models from day one.

Tarifs · Docs