A production-ready Engram for calling OpenAI (and Azure OpenAI) Chat Completions inside bobrapet workflows. It supports rich prompting, function tools, structured outputs, and story dispatch, and runs as either a batch Job or a streaming Deployment.

• OpenAI & Azure OpenAI – Auto-detects Azure when AZURE_ENDPOINT is present, otherwise uses OpenAI’s public API.
• Function tools + structured outputs – Register tools via spec.with.tools; pipe structured arguments back through structuredToolName.
• Story dispatch – Map tool invocations to downstream StoryRuns using dispatchTools.
• Dual mode – Same implementation powers batch step runs and streaming deployments with backpressure.
• Per-request overrides – Customize model, temperature, prompts, or history per invocation without editing configuration.

make lint
go test ./...
make docker-build

Bundle the image in your bobrapet release or run locally with make run (requires bobrapet execution environment variables).

Example:

defaultModel: gpt-4o-mini
defaultTemperature: 0.2
tools:
  - name: run_story
    description: Trigger a follow-up story with inputs
    parameters:
      type: object
      properties:
        storyName: { type: string }
        inputs: { type: object }
structuredToolName: extract_fields
dispatchTools:
  run_story: create-ticket-story

When structuredToolName is set, the model's call to that tool extracts its JSON arguments into Result.Structured instead of executing the tool. This is useful for structured data extraction (e.g., extracting fields from a conversation).

When dispatchTools is configured, a tool call matching a key in the map automatically triggers the mapped Story. The tool's arguments become the StoryRun inputs. If both structuredToolName and dispatchTools reference the same tool, structured extraction takes precedence.

// Example output when structuredToolName = "extract_fields"
{
  "structured": {
    "customer_name": "Alice",
    "issue_type": "billing",
    "priority": "high"
  },
  "text": "I've extracted the relevant fields from the conversation."
}

Secrets are mounted via mountType: env in the EngramTemplate.

Inputs are provided through the StepRun spec.inputs or streaming StreamMessage. Supported fields:

Advanced request-time overrides mirror the template schema. You can override topP, maxTokens, maxCompletionTokens, presencePenalty, frequencyPenalty, stop, modalities, store, parallelToolCalls, logprobs, topLogprobs, seed, serviceTier, reasoningEffort, verbosity, responseFormat, metadata, logitBias, audio, promptCacheKey, safetyIdentifier, user, tools, functions, toolChoice, functionCall, choices, allowedTools, useResponsesAPI, prediction, and webSearch per invocation.

Example payload:

{
  "history": [
    { "role": "user", "content": "Create a ticket for server outage." }
  ],
  "systemPrompt": "Be concise.",
  "userPrompt": "Priority should be high and tags: ops, urgent",
  "temperature": 0.1,
  "model": "gpt-4o-mini"
}

The Engram returns:

Streaming consumes engram.InboundMessage on input and emits engram.StreamMessage on output:

• If Inputs is set, it is decoded into the request payload described above.
• If Inputs is empty, the Engram attempts to decode Payload as the same structure.
• Metadata is copied and enriched with type, provider, and model fields.
• Call msg.Done() after successful handling so ordered/replay-capable transports can advance acknowledgement state.

Responses are emitted as StreamMessage with Payload containing the JSON-encoded output map shown above, and the same bytes mirrored into Binary with MimeType: application/json.

• make lint – GolangCI-Lint using the shared config.
• go test ./... – Compile and run unit tests.
• make run – Execute locally against the bobrapet runtime environment.
• make docker-build – Multi-stage image build (honours IMG).

• Contributing
• Support
• Security Policy
• Code of Conduct
• Discord

Copyright 2025 BubuStack.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.