Local AI language models are extraordinary reasoning engines. They cannot search the web, remember last Tuesday, fetch a box score, or save a file. They are a genius locked in a room with no windows and no doors — answering questions slipped under the door on pieces of paper.
The Phantom opens the windows. It is a Python wrapper around Ollama that intercepts every conversation turn, performs whatever research or data retrieval is needed before the model sees the prompt, injects the results as context, and handles all post-processing after the model responds.
"The model does one thing: it reasons. The Phantom does everything else."
The Phantom performs all of the following before passing a prompt to the local model:
Scrapes full article content (not snippets) via Trafilatura with BeautifulSoup fallback. Attempts paywall bypass via 12ft.io cascade. DuckDuckGo search returning full article text.
MLB Stats API integration (no API key required) for live scores and full season statistics with correct at-bat vs plate appearance distinction. RSS feed reading from configured sources. YouTube transcript fetching without API key.
ChromaDB vector database running locally. Every conversation turn stored as a mathematical vector. Before each turn, the five most semantically similar past turns retrieved — not by keyword, but by meaning. How human memory actually works.
The model reads web content, live data, and relevant past conversations as if it simply knew these things. It did not simply know them. The Phantom knew where to look.
Connects to any local Ollama model at localhost:11434. Maintains conversation history in JSON. Multi-project isolation with separate history, library, and workspace per project. Identity file system for personality and context definition.
Two-stage detection: keyword triggers plus LLM self-query ("do I need the web for this?"). Catches queries that keyword matching would miss — "How has Miguel Rojas performed this season?" has no trigger keyword but the model correctly identifies the need.
phantom_scout.py runs independently on a schedule, fetches fresh data, asks a local model to write content, renders self-refreshing HTML. Powers phantom.fluidfortune.com — live, AI-written content with no cloud dependency and no ongoing cost.
ChromaDB entirely local. Retrieves by semantic similarity, not keyword match. Auto-distillation keeps 20 most recent turns verbatim, summarizes older turns into memory snapshots. Context window stays sharp indefinitely.
The Phantom was originally conceived as a standalone AI assistant. It became Tier 0 of the Pisces Moon ecosystem — the data preparation layer the whole platform was always waiting for. Wardrive CSVs from the T-Deck Plus feed directly into The Phantom's analysis pipeline. The baseball app on the Linux tablet queries The Phantom's MLB endpoint instead of cloud AI. The data formats are the same by design.