Extract (code, description) from a model response.

Emit run-level F1 metadata for Okahu/Monocle.

Emit a small trace span with judge score metadata.

Run evaluation work in a small thread pool.

Select all files of a specific type.

Yield SSE-formatted chunks from the vLLM streaming response.

Final guardrails after vLLM returns text.

Gemma/vLLM sometimes prepends a verbatim copy of the system message before the answer.

Returns (value, extraction_score, citation_dict_or_None).

Returns (value, extraction_score, citation_dict_or_None).

The Contrarian: acts as a dedicated Red Team. Given the Analyst's bull case, it searches specifically for risks, counter-arguments, regulator

Basic resume extraction using regex patterns (fallback)

Create the LangChain agent backed by Nebius ChatNebius.

Compile the Cosmos Arena debate as an explicit LangGraph state machine. Each council member is a node and a single model call. The graph loops th

Build GitHub API URL for repository operations. Args: owner: Repository owner repo: Repository name path: Opti

Create the shared Cosmos reasoner backed by Nebius Token Factory.

Build a very small LangGraph graph with two nodes: - generate_quiz - evaluate (called after the UI has collected answers) The UI will

Generate chat completion using Nebius AI

Complete a human review (synchronous, for Streamlit).

Crawl a single conference with default settings Args: conference_url: URL of the conference schedule batch_size: Number

Create search index for a specific conference

Create index for a specific conference ID

Create indexes for all conferences that are missing them

Create a standardized progress display component.

Factory function to create query retriever.

Create a repository selection dropdown.

Create a status display textbox.

Discover files in a GitHub repository. Args: repo_url: GitHub repository URL or owner/repo format branch: Git branch na

Display assistant message with thinking content if present, and always render the full response.

Display assistant message with thinking content if present.

Run all queued/submitted jobs for the given documents to a terminal state. Returns stats: {succeeded, failed, timed_out}.

Run all queued/submitted jobs for the given documents to a terminal state. Returns stats: {succeeded, failed, timed_out}.

Edit a LinkedIn post based on reviewer feedback. Args: post: The current post to edit. reviews: List of reviews to address.

Create the .memory/ directory if it doesn't exist and return its path.

Escape dollar signs to prevent markdown math mode issues.

Download a small text file with basic error handling.

Fetches filtered cars from MongoDB. Optional: - max_price - city - keyword (brand or any text, e.g. 'bmw')

Find every literal occurrence of `match_value` across all documents via full-text search (this is "replace this value everywhere"). Each occurrenc

Formats the response from a handoff_to_user call for display.

Format post examples as text for prompt injection.

Format SERP results into a readable format for agent analysis.

Convert Decimal to float, handling None.

Generate embedding for a given text. Args: text: Text to embed Returns: List of