Run both agents and return a result dict with keys: response, protocol_title, protocol_content, source, score

Retrieve relevant chunks and generate an answer with Gemma 4.

Run a Google SERP query via Scrapingdog and return organic results.

Embed and insert LangChain Documents into ChromaDB.

Generate a grounded customer support response. Called when the orchestrator routes to the resolution path.

Send a system and user prompt to the Orq.ai router and return the text response.

Convert ScoredPoint results into a standardised context dict.

Embed a single piece of text and return its vector.

Create and cache the Mem0 MemoryClient singleton.

Renders all pages of a PDF (bytes) to a list of PNG image bytes.

Render a platform section header with a sentiment pill and the report body.

Switch the database the agent tools query against.

Ask Gemini to describe/transcribe a media file via its URI.

Return the persistent ChromaDB collection, creating it if needed.

Create a mock for urllib.request.urlopen.

Shared Orq.ai router client used by every agent.

Deterministic regex-only decision used when LLM output cannot be parsed.

Split raw text into chunks separated by blank lines.

Collect the final response text from a single runner.run_async() call.

Accept both list and {'results': [...]} formats from mem0.

Upload a file to Gemini File API, wait for processing, return URI.

Embed `text_key` from each record and upsert into the collection.

Return an alphabetically sorted list of indexed band titles.

Return a ChromaDB-safe collection name derived from the video ID.

Run vision extraction on a preprocessed PIL image. Returns a structured dict with receipt fields.

Fetch and return the text content of a web page. Args: url: URL of the page to fetch. max_chars: Maximum characters to return (de

Return the currently active database path from session state.

Return the cached persistent ChromaDB client.

Convert PIL Image to bytes. Args: img (PIL.Image): Image to convert format (str): Image format (default: 'PNG')

Load every .md and .txt file in protocols_dir into Qdrant. Wipes any existing entries first so re-ingestion is idempotent.

Load all transactions from SQLite into a DataFrame, returning empty if no DB exists.

Return the top-k most relevant chunks for the query using BM25.

Return a Plotly bar or pie chart for the DataFrame, or None if the data is not plottable.

Return the system prompt assigning a debater to its position on the given topic.

Return the user-turn prompt instructing the debater to deliver their argument for this round.

Remove duplicate strings from a list using case-insensitive comparison while preserving order.

Render the debate so far as plain text for the next debater.

Check whether an error message looks like a 429 quota/rate-limit error.

Check whether an error message looks like a temporary model-overload error.

Classify by keyword matching when the model returns non-JSON.

Text-only context for Kimi — never replay assistant tool_calls to the API.

Optimize image for local LLM consumption.

Optimize image for local LLM consumption.

Parse raw SQL query results (JSON list or pipe-delimited string) into a DataFrame.

Re-chunk all band docs and rebuild the BM25 index.

Run the ADK agent and collect the final response text.

Calls Ollama chat endpoint with streaming enabled.

Calls Ollama chat endpoint with streaming enabled.

Call _stream_response, retrying on transient model errors. 503/UNAVAILABLE overload errors use the existing short exponential backoff. 429/RE

Remove markdown ```python ... ``` fences that the model may wrap around code.

Upload a local audio file to the Gemini Files API and return the file object.

Search YouTube via Scrapingdog and return video results.

Fetch and add a band to the knowledge base. Rebuilds the index. Returns the resolved Wikipedia title.

Create a fresh Agno agent with the provided credentials.

Build the shared LLM (OpenAI SDK -> Orq.ai router).

Initialise and return all pipeline components. Call once at startup and reuse across queries. Returns: (qdrant_client, embeddings, orq_cl

Return a 384-dimensional embedding for a single text string.

Return embeddings for a list of texts.

Embed a list of documents (index time).

Embed a list of texts with Mistral Embed, batching the requests.

Escape Markdown special characters for Telegram.

Safely extract a plain string from an LLM response. langchain-google-genai can return response.content as either a str or a list of conte

Fetch articles from all feeds in sources_path published within the last `hours`. Args: sources_path: Path to the JSON file listing RSS fe

Fill PDF widgets using a pre-calculated logical→internal mapping. data: {logical_name: value} mapping: {logical_name: pdf_internal_wid

High-speed In-Memory Stamping Engine. Uses fuzzy matching and coordinate caching for 50ms frame renders.

Format a timestamp in seconds as a human-readable M:SS or H:MM:SS string.

Return all stored memories for the given user, or an empty list on error.