momentum-system/youtube_claims/recap/prompt.py

"""The extraction instructions (EXTRACT_SPEC.md, adapted for a local model) and the JSON
schema handed to Ollama's structured-output mode. The spec file itself lives in the
git-ignored data dir; this is the versioned copy the pipeline actually uses."""

from __future__ import annotations

SETUPS = ["micro_pullback", "bull_flag", "first_pullback", "breakout_whole_dollar",
          "break_of_premarket_high", "halt_resume", "dip_and_curl", "vwap_break", "abcd",
          "red_to_green", "other"]
KINDS = ["daily_recap", "watchlist", "education", "challenge_summary", "other"]

SYSTEM = """You extract a structured record from ONE transcript of Ross Cameron's (Warrior Trading)
day-trading video. Record what he SAID, not what you think is right.

QUOTES ARE CHECKED BY A PROGRAM. Every `quote`, `setup_quote`, `catalyst_quote` and `market_quote`
must be copied word-for-word from the TRANSCRIPT (a contiguous run of 6-30 words, ASR noise
included; do not fix grammar or numbers). Anything whose quote is not found verbatim is deleted.
If you cannot quote it, leave the item out.

The transcript is speech-to-text: tickers and prices are often misheard. Put the ticker AS HEARD in
`ticker_heard` and your best guess in `ticker_guess`, preferring a symbol from OUR SCANNER LIST when
it sounds alike and the price fits. Spoken prices: "550" usually means $5.50, "750" means $7.50,
"a dollar twenty" = 1.20, "four and a half" = 4.50. Prices are verified against the tape later, so
only give a price he actually said.

What goes where:
- video_kind: daily_recap (he recaps his own trading of a session -- even if he mentions a challenge
  or account milestone), watchlist (mainly stocks to watch for the next session), challenge_summary
  (ONLY when the video summarises a multi-day/multi-week challenge as a whole), education (a lesson
  or story not about one session), other.
- session_date: the trading day described (usually the upload date; the prior Friday for a weekend
  watchlist; null if no specific day).
- trades: walk through the transcript IN ORDER and create ONE entry every time he says he bought,
  got in, took a trade, added, or got back in ("I got in", "I bought", "I was in", "I added",
  "I took a trade on", "I jumped in"). Several trades on the same stock are normal -- each re-entry
  or add is its own entry (add_on=true). Trades he took but mentions only briefly still count, and
  watchlist or challenge videos can contain trades too. Do NOT create trades for stocks he only
  watched, showed on a chart, or talks about hypothetically.
  setup: one of micro_pullback, bull_flag, first_pullback, breakout_whole_dollar,
  break_of_premarket_high, halt_resume, dip_and_curl, vwap_break, abcd, red_to_green, other.
  outcome: win / loss / scratch as he describes it, else null. pnl/shares only if stated.
- passed_on: EVERY other stock he names that he did NOT trade -- top gainers he skipped, stocks on
  his watchlist, names he says he missed or avoided -- with the reason (or what he said about it) and
  a quote naming or describing it. Usually 3-10 per video.
- ticker_guess: when the heard name is close to a symbol in OUR SCANNER LIST (sound-alike, spelled
  letters, or the price he mentions fits), use that scanner symbol.
- rules_stated: general principles he states (risk, sizing, stock selection, when to stop,
  statistics about his accuracy or P&L), each with a quote.
- day.market_temperature: hot (lots of big clean movers, strong momentum), cold (slow, few good
  setups, choppy, tough month), mixed, or null if he gives no read. market_quote: the sentence where
  he characterises the market.
- day.net_pnl: the day's $ result as stated (negative for red); green = true if the day is positive.

Keep paraphrases (reason, rule, lesson, read) short. Skip sponsor segments, course promotion and
filler. Return ONLY the JSON object."""

USER_TMPL = """{part_note}Upload metadata and the day's context follow, then the transcript.

{bundle}"""

PART_NOTE = ("This is PART {i} of {n} of one transcript (parts overlap slightly). Extract only what "
             "is said in this part; fill `day` fields only if this part states them.\n\n")


def _nstr():
    return {"type": ["string", "null"]}


def _nnum():
    return {"type": ["number", "null"]}


def _quoted(key: str) -> dict:
    return {"type": "array", "items": {"type": "object", "properties": {
        key: {"type": "string"}, "quote": {"type": "string"}}, "required": [key, "quote"]}}


SCHEMA = {
    "type": "object",
    "properties": {
        "video_kind": {"type": "string", "enum": KINDS},
        "session_date": _nstr(),
        "day": {"type": "object", "properties": {
            "net_pnl": _nnum(), "green": {"type": ["boolean", "null"]},
            "market_temperature": {"type": ["string", "null"], "enum": ["hot", "cold", "mixed", None]},
            "market_quote": _nstr(), "max_loss_hit": {"type": ["boolean", "null"]},
            "daily_goal_comment": _nstr(), "sizing_comment": _nstr()},
            "required": ["net_pnl", "green", "market_temperature", "market_quote"]},
        "trades": {"type": "array", "items": {"type": "object", "properties": {
            "ticker_heard": {"type": "string"}, "ticker_guess": {"type": "string"},
            "entry_time_hint": _nstr(),
            "setup": {"type": "string", "enum": SETUPS}, "setup_quote": {"type": "string"},
            "entry_price": _nnum(), "exit_price": _nnum(), "stated_high": _nnum(), "stated_low": _nnum(),
            "shares": _nnum(), "pnl": _nnum(),
            "outcome": {"type": ["string", "null"], "enum": ["win", "loss", "scratch", None]},
            "add_on": {"type": "boolean"},
            "stock_context": {"type": "object", "properties": {
                "catalyst": _nstr(), "catalyst_quote": _nstr(), "float": _nstr(),
                "pct_up_at_entry": _nstr(), "leading_gainer_rank": _nnum(),
                "other": {"type": "array", "items": {"type": "string"}}}},
            "why_entered": _quoted("reason"), "tape_reads": _quoted("read"),
            "why_exited": _quoted("reason"), "mistake_or_lesson": _quoted("lesson")},
            "required": ["ticker_heard", "ticker_guess", "setup", "setup_quote", "entry_price",
                         "exit_price", "outcome", "why_entered"]}},
        "passed_on": {"type": "array", "items": {"type": "object", "properties": {
            "ticker_heard": {"type": "string"}, "ticker_guess": {"type": "string"},
            "reason": {"type": "string"}, "quote": {"type": "string"}},
            "required": ["ticker_heard", "ticker_guess", "reason", "quote"]}},
        "rules_stated": _quoted("rule"),
    },
    "required": ["video_kind", "session_date", "day", "trades", "passed_on", "rules_stated"],
}


_TASK_A = ("TASK FOR THIS CALL: your main job is `trades` -- list EVERY position he took, in order, "
           "with its quotes and prices. Also fill video_kind, session_date and day (net P&L, "
           "market_temperature). Passed-on stocks and rules are extracted in a separate call.\n\n")
_TASK_B = ("TASK FOR THIS CALL: fill passed_on (every stock he names but did NOT trade) and "
           "rules_stated (every general principle, statistic or lesson he states -- usually 5-12). "
           "Trades are extracted in a separate call.\n\n")
SCHEMA_A = {"type": "object", "properties": {k: SCHEMA["properties"][k] for k in
                                             ("video_kind", "session_date", "day", "trades")},
            "required": ["video_kind", "session_date", "day", "trades"]}
SCHEMA_B = {"type": "object", "properties": {k: SCHEMA["properties"][k] for k in ("passed_on", "rules_stated")},
            "required": ["passed_on", "rules_stated"]}
SPLIT_TASKS = [(_TASK_A, SCHEMA_A), (_TASK_B, SCHEMA_B)]