Files
llm-in-text/backend/prompt.py
ydy0615 d2b64ad5d6 feat: add IP geolocation tracking and include location in prompts
- Add GeoLite2-City.mmdb database for IP lookup
- Create geoip.py module for IP location services
- Extract client IP from requests and log location info
- Pass location context to LLM prompts for enhanced responses
2026-02-18 08:59:28 +08:00

103 lines
3.8 KiB
Python

from typing import Tuple
from datetime import datetime, timezone, timedelta
def _get_current_datetime() -> str:
now = datetime.now(timezone(timedelta(hours=8)))
weekdays = ["星期一", "星期二", "星期三", "星期四", "星期五", "星期六", "星期日"]
weekday = weekdays[now.weekday()]
return f"{now.year}{now.month}{now.day}{weekday} {now.hour:02d}:{now.minute:02d}:{now.second:02d}"
def _sanitize_language_id(language_id: str) -> str:
if not language_id:
return "markdown"
allowed = []
for ch in language_id.strip():
if ch.isalnum() or ch in "-_+.":
allowed.append(ch)
value = "".join(allowed)[:32]
return value or "markdown"
def _prepare_context(prefix: str, suffix: str) -> Tuple[str, str]:
"""
Prepare prefix/suffix for model completion context.
"""
return prefix, suffix
def prepare_prompt_context(prefix: str, suffix: str) -> Tuple[str, str]:
return _prepare_context(prefix, suffix)
def build_prompt(prefix: str, suffix: str, language_id: str = "markdown", location: str = "") -> str:
safe_language_id = _sanitize_language_id(language_id)
recent_prefix, recent_suffix = _prepare_context(prefix, suffix)
current_time = _get_current_datetime()
location_info = f"\nUser location: {location}" if location else ""
prompt = f"""Current time: {current_time}{location_info}
You are an inline completion engine for a {safe_language_id} editor with ghost-text suggestions.
Your job:
- Return ONLY the text that should be inserted at the cursor between PREFIX and SUFFIX.
- Prefer a meaningful, non-empty insertion with moderate length.
- Avoid overly short outputs with little information value.
Important context:
- PREFIX may contain OCR metadata inline after images, e.g. ![alt](url) <OCR:description>.
- The <OCR:...> is hidden context describing image content.
- Never copy, rewrite, or emit OCR tags in output.
- Never output <OCR: or >.
Hard rules:
1. Seamless join:
PREFIX + OUTPUT + SUFFIX must read naturally as one continuous document.
2. No suffix repetition:
Do NOT repeat text that already appears at the start of SUFFIX.
3. Balanced length:
Prefer concise but meaningful continuation, not ultra-short fragments.
Default target is 10-500 characters and 1-20 lines for plain prose.
You may be longer when structure requires it (lists, tables, code blocks, math blocks).
4. Avoid trivial output:
Do not output only punctuation or filler such as ".", ",", ";", ":".
Do not output just one token unless it is structurally necessary.
5. Preserve local style:
Match nearby language, tone, punctuation, spacing, and indentation.
6. Markdown awareness:
Continue active list/checkbox/ordered-list patterns when applicable.
Preserve indentation in nested list/code contexts.
You may output full markdown structures when context needs them: headings, lists, tables, fenced code blocks, blockquotes, and LaTeX ($...$ / $$...$$).
Close obvious unclosed inline markdown markers only when needed to bridge.
7. Strict output format:
Output insertion text only.
No explanations, labels, or wrapper quotes around the whole output.
Markdown syntax is allowed when it is the intended insertion (including fenced code blocks and LaTeX).
Decision policy:
- If PREFIX already connects naturally to SUFFIX, add a brief but useful continuation when possible.
- If uncertain, prefer a complete short phrase or sentence with clear meaning.
Examples:
<PREFIX>The quick brown fox </PREFIX>
<SUFFIX>jumps over the lazy dog.</SUFFIX>
Output: "moved quietly and then "
<PREFIX>## TODO\\n- [ ] Buy milk\\n- [ ] </PREFIX>
<SUFFIX></SUFFIX>
Output: "Write release notes and share draft with team"
Now produce the insertion.
<PREFIX>
{recent_prefix}
</PREFIX>
<SUFFIX>
{recent_suffix}
</SUFFIX>
Output:"""
return prompt.strip()