Return a connected VectorAI DB client.

Get embedding for text or image using Cohere Embed-4. Args: api_key (str): Cohere API key input_data (str or bytes): Tex

Return the cached SentenceTransformer embedding function for ChromaDB.

Return existing RAG instance or create a new one.

Embed N hypothetical documents and average their vectors. The averaged vector sits closer in embedding space to real relevant documen

Return the total number of chunks currently in the BM25 index.

Parse CSV, categorize transactions, and persist to SQLite.

Create tables and seed sample data if the database doesn't exist yet.

Return expenses as a DataFrame, optionally filtered.

Extract the display text from a memory object returned by Mem0.

Render agent intermediate steps inside a Streamlit expander.

Run the browser-use agent for a single natural language instruction. Args: instruction: The user's natural language task. max_ste

Translate the strategy into a full creative campaign plan.

Research the market using Serper search and return a structured brief.

Turn market research into a full marketing strategy.

Create all four collections if they do not already exist.

Split a streamed buffer into (reasoning, answer) for the two panels.

Persist a resolved query into VectorAI DB memory. This collection grows over time and is searched alongside the static sources.

Render the result as a chat-friendly markdown block.

Search the web using DuckDuckGo Instant Answer API. Returns a JSON string with top results including titles, URLs, and snippets. Args:

Build a YouTube URL that starts playback at the given second offset.

Run an agent once (no streaming) and return its text content.

Build and return an OpenAI client pointed at the Orq.ai router.

Build the contents list for a multi-turn conversation. history is a list of {"role": "user"|"assistant", "content": "..."} dicts. The aud

Narrow prompt for the LLM: produce only a short empathetic acknowledgment. No context documents -- the LLM does not need them and should not

Build and return the ChatOpenAI client pointed at the Orq.ai router.

Build the LangChain SQL toolkit and replace the default query tool with the read-only version.

Generate a deterministic case reference from the current Unix timestamp.

Route a PDF page to text or vision extraction based on the extracted character count.

Remove Wikipedia markup artifacts.

Derive display column names from the SELECT list when results are bare tuples.

Build and compile the full LangGraph SQL agent workflow against the Chinook database.

Deduplicate lab findings by test name, keeping the most severe status.

Return an HTML pill based on keywords in the report.

Assemble the final escalation message. The empathy paragraph comes from the LLM. All structural facts are hardcoded here and never left to LL

Return a short text snippet from the entry body. Prefers the full content field, falls back to summary/description. Strips to a max of 500 ch

Best-effort extraction of useful fields from a browser-use history object. The browser-use API surface has shifted across versions, so each acces

Extract text from a PDF and split into page-level chunks.

Extract the last SQL query and its raw result string from the agent message history.

Gradio may store message content as a plain string or a list of content blocks.

Format the list of flagged abnormal values into a human-readable string.

Build and return an OpenAI client pointed at the Orq.ai router.

Return the appropriate device and compute type based on CUDA availability.

Lazily build the Mistral client so import never fails without a key.

Return the cached KPipeline for the given Kokoro language code, creating it if needed.

Fetch Google News results for a query via Scrapingdog.

Return True if the knowledge directory has at least one real file.

Convert raw image bytes to a vision ProcessedPage, normalizing to RGB PNG.

Renders each page of a PDF to a PNG byte string, up to max_pages.

Returns True if the text appears to contain an HTML table.

Merge two PatientInfo objects, preferring the first non-empty value for each field.

Return a list of required environment variable names that are not currently set.

Extract a timezone-aware datetime from a feedparser entry. Tries published_parsed first, then updated_parsed. Returns None if neither is avai

Extract JSON from the model response robustly. Handles cases where the model wraps the output in markdown fences.

Extract and parse the first JSON object found in an LLM response string.

Strip markdown fences from the LLM response and parse it into a ClinicalProfile.

Best-effort extraction of the judge's JSON, with graceful fallbacks.

Run the full regex pass and return a structured summary dict.

Format the running transcript as markdown.