PromptCaller

PromptCaller centralizes .prompt files in one folder, turns prompt responses into structured outputs, and includes SKILLS so agents can create and maintain prompt files consistently.

Features

• Load prompts from a .prompt file containing a YAML configuration and a message template.
• Invoke prompts using LangChain with OpenAI, Google Gemini, or Ollama, with support for structured output and tools.

Installation

To install the package, simply run:

pip install prompt-caller

For OpenAI prompts, you will also need an .env file that contains your OpenAI API key:

OPENAI_API_KEY=your_openai_api_key_here

Model Providers

Choose a provider in the .prompt frontmatter:

When provider is omitted, model names beginning with gemini select Google; all other names select OpenAI. An explicit provider takes precedence over the model name. provider, output, and types are internal PromptCaller metadata; the remaining frontmatter configures the selected chat client. Use parameters supported by that client (for example, reasoning for Ollama).

Local Ollama

Install and start Ollama, then download the course model:

ollama pull qwen3.5:4b-q4_K_M

Use prompts/sample-ollama.prompt:

---
provider: ollama
model: qwen3.5:4b-q4_K_M
base_url: http://localhost:11434
temperature: 0
reasoning: false
output:
  answer: "string | Answer to the question"
  confidence: "enum[low|medium|high] | Confidence level"
---
<system>
Answer the question clearly and concisely.
Return all requested structured fields.
</system>

<user>
{{question}}
</user>

from prompt_caller import PromptCaller

ai = PromptCaller()
response = ai.call("sample-ollama", {"question": "What is 7 + 5?"})
print(response.answer)

Local Ollama does not require an OpenAI key. base_url is optional; when omitted, the Ollama client uses its default server address. Tool-enabled agents require a model that supports tool calling. PromptCaller uses the standard LangChain structured-output and tool mechanisms.

For Ollama agents with both tools and an output schema, PromptCaller runs the tool-enabled agent without a response schema, then makes one additional call to the same model with with_structured_output(). The formatting call receives the system prompt and full conversation, including tool results, without bound tools. Parsing and validation errors propagate rather than falling back to plain text. This avoids constraining tool calls with Ollama's JSON schema format. OpenAI and Gemini keep their existing agent response strategy; other Ollama paths are unchanged. allowed_steps limits the agent graph, not the additional formatting call.

Real tests with qwen3.5:4b-q4_K_M passed for structured call() output and agent() with actual tool execution, both with and without an output schema.

CLI Skill Installation

PromptCaller ships with a CLI command to install a PromptCaller skill pack into .agents/skills.

prompt-caller install

Alternative module invocation:

python -m prompt_caller install

By default, this installs to .agents/skills/prompt-caller and overwrites existing files.

Usage

1. Define a prompt file:

Create a .prompt file in the prompts directory, e.g., prompts/sample.prompt:

---
model: gpt-5.2
reasoning_effort: medium
output:
  result: "Final result of the expression"
  explanation: "Explanation of the calculation"
---
<system>
You are a helpful assistant.
</system>

<user>
How much is {{expression}}?
</user>

This .prompt file contains:

• A YAML-like header for configuring the model and parameters.
• A template body using Jinja2 to inject the context (like {{ expression }}).
• Messages structured in a JSX-like format (<system>, <user>).

2. Load and call a prompt:

from prompt_caller import PromptCaller

ai = PromptCaller()

response = ai.call("sample", {"expression": "3+8/9"})

print(response)

In this example:

• The expression value 3+8/9 is injected into the user message.
• The model will respond with both the result of the expression and an explanation, as specified in the output section of the prompt.

Advanced Prompt Example (sample-5.2-complete.prompt)

Use this when you want strongly typed, multi-field structured output with reusable types:

---
model: gpt-5.2
reasoning_effort: high
output:
  result: "number | Final result of the expression"
  explanation: "string | Explanation of the calculation"
  steps: "list[Step] | Ordered calculation steps"
  confidence: "enum[low|medium|high] | Confidence level for the computed answer."

types:
  Step:
    expression: "string | Expression evaluated in this step"
    value: "number | Numeric result of this step"
---
<system>
  You are a helpful assistant and you have access to tools.
  Use tools when needed.
  Return all requested structured fields.
</system>

<user>
  How much is {{expression}}?
</user>

Example call:

from prompt_caller import PromptCaller

ai = PromptCaller()

response = ai.call("sample-5.2-complete", {"expression": "(3 + 8) / 9"})
print(response)

3. Using the agent feature:

The agent method allows you to enhance the prompt's functionality by integrating external tools. Here's an example where we evaluate a mathematical expression using Python�s eval in a safe execution environment:

from prompt_caller import PromptCaller

ai = PromptCaller()

def evaluate_expression(expression: str):
      """
      Evaluate a math expression using eval.
      """
      safe_globals = {"__builtins__": None}
      return eval(expression, safe_globals, {})

response = ai.agent(
      "sample-agent", {"expression": "3+8/9"}, tools=[evaluate_expression]
)

print(response)

In this example:

• The agent method is used to process the prompt while integrating external tools.
• The evaluate_expression function evaluates the mathematical expression securely.
• The response includes the processed result based on the prompt and tool execution.

Response Metadata

Both call() and agent() accept the keyword-only option include_metadata=False. By default their return values and error behavior are unchanged. Set it to True to attach a reserved _metadata attribute to the returned object, without a wrapper:

response = ai.call(
    "sample-5.2-complete",
    {"expression": "(3 + 8) / 9"},
    include_metadata=True,
)
print(response.result)  # Same access as before
print(response._metadata["response_id"])
print(response._metadata["response_metadata"])
print(response._metadata["usage_metadata"])

response = ai.agent(
    "sample-agent",
    {"expression": "3+8/9"},
    tools=[evaluate_expression],
    include_metadata=True,
)
print(response._metadata)

The attribute always contains these three keys:

The existing object type, prompt schema, model_dump(), and model_dump_json() are preserved. When enabled, metadata is attached to a shallow model copy so cached responses remain untouched; object identity is not preserved in this mode. _metadata is not included in those serialized results. Include it explicitly when constructing a payload if you need to persist the metadata. The option belongs in the Python call, not the prompt frontmatter, and does not add model requests or switch provider APIs. IDs are not necessarily resp_...; for example, OpenAI Chat Completions can return chatcmpl-... IDs.

For ordinary agents, the source is the last AIMessage, even when followed by a tool message. If the last AI message has no provider ID, response_id is None; an earlier response's ID is not substituted. For Ollama agents with tools and a schema, the source is the final structured-formatting call. These values do not represent accumulated usage across agent calls, tools, or retries.

Metadata attachment supports Pydantic BaseModel outputs, including AIMessage. When using agent(output=...) with this option, pass a Pydantic model class. Other schema formats (such as TypedDict, dataclasses, JSON Schema dictionaries, and explicit LangChain strategy objects) are rejected before the agent runs; their existing behavior is unchanged when the option is disabled. _metadata is reserved: conflicting prompt fields, private attributes, properties, or existing instance attributes raise ValueError rather than being overwritten. Known schema conflicts are checked before model invocation. Unexpected return types raise TypeError. Parsing and provider errors continue to propagate.

How It Works

1. _loadPrompt: Loads the prompt file, splits the YAML header from the body, and parses them.
2. _renderTemplate: Uses the Jinja2 template engine to render the body with the provided context.
3. _parseJSXBody: Parses the message body written in JSX-like tags to extract system and user messages.
4. call: Invokes the selected chat provider with the parsed configuration and messages, and handles structured output via dynamic Pydantic models.

Build and Upload

To build the distribution and upload it to a package repository like PyPI, follow these steps:

1. Build the distribution:

   Run the following command to create both source (sdist) and wheel (bdist_wheel) distributions:

   python setup.py sdist bdist_wheel
   

   This will generate the distribution files in the dist/ directory.

2. Upload to PyPI using Twine:

   Use twine to securely upload the distribution to PyPI:

   twine upload dist/*
   

   Ensure you have configured your PyPI credentials before running this command. You can find more information on configuring credentials in the Twine documentation.

Tests

pytest --cov=prompt_caller ; coverage report --sort=miss

The regular suite does not contact Ollama. To run the real course-model tests:

$env:RUN_OLLAMA_TESTS = "1"
python -m pytest tests/test_ollama_integration.py -v

On Bash, use RUN_OLLAMA_TESTS=1 python -m pytest tests/test_ollama_integration.py -v. These tests use exactly qwen3.5:4b-q4_K_M and cover plain calls, typed output with named types, and actual tool execution with and without structured agent output, including an explicit output= override. Tool codes are generated randomly so the agent must obtain them from the tool rather than infer them from the prompt. Each real-model scenario runs with response metadata disabled and enabled. Set OLLAMA_BASE_URL to use another server. When enabled, the tests fail with a clear error if the server is unavailable or the model has not been downloaded.

Output Schema DSL

The output field supports both legacy and typed schema definitions.

Legacy format (defaults to string):

output:
  answer: "Final answer to return"

Compact DSL format (recommended):

output:
  title: "string | Final title"
  confidence: "enum[low|medium|high] | Confidence level"
  steps: "list[Step] | Ordered calculation steps"
  note?: "string | Optional extra note"

Supported type expressions:

• string
• number
• integer
• boolean
• list[T]
• enum[a|b|c]
• named type references declared under top-level types

Named reusable types:

types:
  Step:
    expression: "string | Expression evaluated in this step"
    value: "number | Numeric result of this step"

Rules:

• Optional fields are declared with ? suffix (for example note?).
• call() uses prompt output when present.
• agent() uses prompt output only when output= is not explicitly passed.

License

This project is licensed under the Apache License 2.0. You may use, modify, and distribute this software as long as you provide proper attribution and include the full text of the license in any distributed copies or derivative works.