JEVANY / DOCUMENTATION

Integrations

Choose a tool

The tool-routing example maps a support request to one of four tools, using the same Python client for local and HTTP inference. With a JevAny server running:

python -m examples.tool_routing "Where is order A-104?" --min-probability 0.8

Use --checkpoint runs/my-jev --device cpu to load a local model instead. The result includes the selected option's probability and the full decision. It returns a tool name only when that probability reaches the threshold; otherwise, "tool": null and "deferred": true let your application ask for more information or hand the case to another decision process.

The example prints a routing decision. Your application supplies the tool arguments and executes the tool. Set the threshold on held-out data from your own task; 0.8 is an example setting. Use the probability distribution for this threshold, since the API's choice confidence adjusts for chance agreement.

Compile questions and decision trees

Jev-Harness uses an external LLM to compile tasks into typed questions. The planner sees an evidence schema by default. The caller retains control of the state, which passes unchanged to JevAny for the bounded decision.

Planner output must be a JSON object. Markdown fences are accepted when their body contains only JSON; place explanations outside the fence.

Local Chat Completions planner

ChatCompletionsGenerator connects to a non-streaming text Chat Completions endpoint. It needs only the base JevAny installation. With a planner serving my-planner at http://127.0.0.1:8000/v1 and a JevAny server at http://127.0.0.1:8008:

python -m examples.harness --planner-model my-planner \
  --task "Choose an execution mode and decide whether rollback is required." \
  --evidence '{"environment":"staging","tests":"passed","snapshot":"available"}'

Use --planner-url to change the planner's API prefix, and --base-url or --checkpoint to choose the JevAny backend. The planner must serve the model name you pass. Local unauthenticated servers need no key; authenticated planners use the optional JEVANY_PLANNER_API_KEY environment variable. Non-loopback endpoints require HTTPS.

The adapter supports max_output_tokens, temperature, top_p, and stop_sequences; the served model must support the parameters you use. It maps the token limit to max_completion_tokens and stop sequences to stop. Responses return text, provider usage, and stop_reason. HTTP errors include the server's explanation; requests are not retried automatically.

Your own planner can implement generate(prompt, params=None) with the same result fields and be passed to JevHarness. The command-line example keeps evidence values with JevAny unless --planner-sees-evidence is enabled. This default is not a prompt-injection security boundary.

Bedrock planner

With the JevAny server running, install the Bedrock adapter:

python -m pip install -e '.[bedrock]'

Use standard AWS credentials with access to the chosen Bedrock model. This example uses the same planner model as the recorded harness run:

from jevany.bedrock import BedrockGenerator
from jevany.harness import HTTPDecisionClient, JevHarness

planner = BedrockGenerator("us.anthropic.claude-opus-4-7")
harness = JevHarness(planner, HTTPDecisionClient("http://127.0.0.1:8008"))
result = harness.run(
    "Choose an execution mode and decide whether rollback is required.",
    {"environment": "staging", "tests": "passed", "snapshot": "available"},
)
print(result["decision"]["answers"])

Jev-Symbolic asks an LLM to write a compact decision tree, validates branch coverage and acyclicity, then sends each internal node to JevAny. Every result records the outcome ID and complete branch trace.

For command-line entry points, see the harness example and the symbolic example. The saved harness and symbolic outputs include the compiled requests and decisions.