Ollama Completions

The news-sentiment-ai-trader system utilizes Ollama-compatible models to perform structured sentiment analysis. The system implements two distinct completion strategies to ensure robust JSON output from LLMs: OllamaOutlineToolCompletion (function calling) and OllamaOutlineFormatCompletion (native JSON mode). Both implementations are built on top of the agent-swarm-kit and target the minimax-m2.7:cloud model.

Model Configuration

All completions are routed through a central configuration that connects to an Ollama host.

• Model Name: minimax-m2.7:cloud,]
• Host: https://ollama.com]
• Authentication: Uses the OLLAMA_TOKEN environment variable]
• Singleton Pattern: The getOllama function uses singleshot to ensure only one instance of the Ollama client is created]

Reliability Parameters

Both completion types share identical retry and timeout logic to handle network instability or model hallucinations:

• COMPLETION_MAX_ATTEMPTS: 3 (Internal loop for model corrections)]
• COMPLETION_MAX_RETRIES: 5 (External retry wrapper for network/timeout errors)]
• COMPLETION_RETRY_DELAY: 5,000ms]
• COMPLETION_TIMEOUT: 300,000ms (5 minutes)]

---

Implementation Strategies

1. OllamaOutlineToolCompletion

This implementation forces the model to use a specific function called provide_answer. It is registered under the name ollama_outline_tool_completion].

Mechanism:

1. Tool Definition: A tool of type function is created with the name provide_answer. The JSON schema for the response is passed into the parameters field].
2. System Prompting: A system message is injected at the start of the conversation, explicitly commanding the model to use the tool and forbidding plain text responses].
3. Fallback/Correction: If the model fails to call the tool, the system appends a user message reminding the model to use provide_answer and increments the attempt counter].
4. Data Repair: The tool arguments are processed through jsonrepair before parsing to fix minor syntax errors in the LLM's JSON string].

2. OllamaOutlineFormatCompletion

This implementation uses Ollama's native structured output capability. It is registered under the name ollama_outline_format_completion].

Mechanism:

1. Schema Injection: The JSON schema is passed directly to the format parameter of the ollama.chat call].
2. Response Handling: The model's message.content is expected to be a raw JSON string.
3. Validation: The parsed JSON is validated against the schema using validateToolArguments from agent-swarm-kit].

Completion Logic Flow

The following diagram illustrates how fetchCompletion handles the request lifecycle for both implementations.

Completion Request Lifecycle

---

Technical Details

JSON Repair and Validation

Because LLMs occasionally output invalid JSON (e.g., trailing commas, missing quotes), the system uses the jsonrepair library. This is applied to both tool arguments and format-mode content before JSON.parse is invoked,].

Timeout Handling

To prevent the trading pipeline from hanging indefinitely, a Promise.race is used between the Ollama request and a sleep timer]. If the timer wins, a COMPLETION_TIMEOUT_SYMBOL is returned, triggering a retry].

Data Flow: From Request to IOutlineMessage

The following diagram maps the transformation of data from the initial call to the final structured message.

Data Transformation Pipeline

Shared Configuration Table

Feature	OllamaOutlineToolCompletion	OllamaOutlineFormatCompletion
Registration Name	ollama_outline_tool_completion	ollama_outline_format_completion
Model	minimax-m2.7:cloud	minimax-m2.7:cloud
Primary Method	Function calling (provide_answer)	Native JSON format schema
Internal Retries	3 attempts with system prompt injection	3 attempts with schema re-validation
Global Retries	5 retries (5s delay)	5 retries (5s delay)
Flags	Russian Language, Reasoning: high	Russian Language, Reasoning: high