Add BLIGHT:: document-scope triggers and case-insensitive matching

- Split TRIGGER_PATTERN into INLINE_PATTERN (BLIGHT:) and DOCUMENT_PATTERN
  (BLIGHT::), both case-insensitive
- Inline triggers replace only the trigger line (existing behaviour)
- Document-scope triggers replace the entire file; multiple BLIGHT:: triggers
  in one file are processed sequentially, each seeing the previous result
- Updated FAILED_TEMPLATE to two-line format with BLIGHT_FAILED and BLIGHT_ERROR
- Added complete_document() to AIProvider ABC and GeminiProvider with a
  dedicated system prompt instructing the model to return the full document

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-03-16 12:36:32 -05:00
parent 90d7089886
commit cf71fb4464
3 changed files with 98 additions and 22 deletions

View File

@@ -4,13 +4,14 @@ from abc import ABC, abstractmethod
class AIProvider(ABC): class AIProvider(ABC):
"""Base class for all AI provider implementations. """Base class for all AI provider implementations.
To add a new provider, subclass this and implement `complete`, then To add a new provider, subclass this and implement `complete` and
instantiate your provider in `processor.py` instead of GeminiProvider. `complete_document`, then instantiate your provider in `processor.py`
instead of GeminiProvider.
""" """
@abstractmethod @abstractmethod
def complete(self, document: str, instruction: str) -> str: def complete(self, document: str, instruction: str) -> str:
"""Process an instruction in the context of a full document. """Process an inline instruction in the context of a full document.
Args: Args:
document: The full markdown document text (for context). document: The full markdown document text (for context).
@@ -19,3 +20,15 @@ class AIProvider(ABC):
Returns: Returns:
The text to insert in place of the trigger line. The text to insert in place of the trigger line.
""" """
@abstractmethod
def complete_document(self, document: str, instruction: str) -> str:
"""Apply a document-scope instruction and return the full rewritten document.
Args:
document: The full markdown document text.
instruction: The BLIGHT:: instruction extracted from the trigger line.
Returns:
The full rewritten document as a string.
"""

View File

@@ -2,7 +2,7 @@ import google.generativeai as genai
import config import config
from .base import AIProvider from .base import AIProvider
_SYSTEM_PROMPT = ( _INLINE_SYSTEM_PROMPT = (
"You are an inline document assistant. " "You are an inline document assistant. "
"The user will provide a markdown document and a specific instruction. " "The user will provide a markdown document and a specific instruction. "
"Your response must contain ONLY the text to be inserted into the document — " "Your response must contain ONLY the text to be inserted into the document — "
@@ -11,13 +11,26 @@ _SYSTEM_PROMPT = (
"Respond as if your output will be dropped directly into the middle of a document." "Respond as if your output will be dropped directly into the middle of a document."
) )
_DOCUMENT_SYSTEM_PROMPT = (
"You are a document editing assistant. "
"The user will provide a markdown document and a specific instruction. "
"Apply the instruction to the entire document and return the full rewritten document. "
"Your response must contain ONLY the rewritten document — "
"no preamble, no explanation, no meta-commentary, no markdown code fences. "
"Preserve the document's structure and formatting unless the instruction says otherwise."
)
class GeminiProvider(AIProvider): class GeminiProvider(AIProvider):
def __init__(self) -> None: def __init__(self) -> None:
genai.configure(api_key=config.GEMINI_API_KEY) genai.configure(api_key=config.GEMINI_API_KEY)
self._model = genai.GenerativeModel( self._inline_model = genai.GenerativeModel(
model_name="gemini-2.5-flash-lite", model_name="gemini-2.5-flash-lite",
system_instruction=_SYSTEM_PROMPT, system_instruction=_INLINE_SYSTEM_PROMPT,
)
self._document_model = genai.GenerativeModel(
model_name="gemini-2.5-flash-lite",
system_instruction=_DOCUMENT_SYSTEM_PROMPT,
) )
def complete(self, document: str, instruction: str) -> str: def complete(self, document: str, instruction: str) -> str:
@@ -25,5 +38,13 @@ class GeminiProvider(AIProvider):
f"DOCUMENT:\n\n{document}\n\n" f"DOCUMENT:\n\n{document}\n\n"
f"INSTRUCTION: {instruction}" f"INSTRUCTION: {instruction}"
) )
response = self._model.generate_content(prompt) response = self._inline_model.generate_content(prompt)
return response.text.strip()
def complete_document(self, document: str, instruction: str) -> str:
prompt = (
f"DOCUMENT:\n\n{document}\n\n"
f"INSTRUCTION: {instruction}"
)
response = self._document_model.generate_content(prompt)
return response.text.strip() return response.text.strip()

View File

@@ -5,8 +5,12 @@ from ai import GeminiProvider
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
TRIGGER_PATTERN = re.compile(r"^BLIGHT:\s+(.+)$", re.MULTILINE) # Inline trigger: BLIGHT: <instruction> (single colon, case-insensitive)
FAILED_TEMPLATE = "<!-- BLIGHT_FAILED: {instruction} -->" INLINE_PATTERN = re.compile(r"^BLIGHT:(?!:)\s+(.+)$", re.MULTILINE | re.IGNORECASE)
# Document-scope trigger: BLIGHT:: <instruction> (double colon, case-insensitive)
DOCUMENT_PATTERN = re.compile(r"^BLIGHT::\s+(.+)$", re.MULTILINE | re.IGNORECASE)
FAILED_TEMPLATE = "<!-- BLIGHT_FAILED: {instruction} -->\n<!-- BLIGHT_ERROR: {error} -->"
_MAX_RETRIES = 3 _MAX_RETRIES = 3
_RETRY_DELAYS = [1, 2, 4] # seconds between attempts _RETRY_DELAYS = [1, 2, 4] # seconds between attempts
@@ -17,42 +21,80 @@ _provider = GeminiProvider()
def process_document(content: str) -> tuple[str, bool]: def process_document(content: str) -> tuple[str, bool]:
"""Scan content for BLIGHT triggers and process each one. """Scan content for BLIGHT triggers and process each one.
Inline triggers (BLIGHT:) are processed first in document order, each
replacing only the trigger line. Document-scope triggers (BLIGHT::) are
processed next in document order, each replacing the entire file content
and operating on the result of the previous.
Returns: Returns:
(updated_content, changed) where changed is True if any triggers (updated_content, changed) where changed is True if any triggers
were found and the content was modified. were found and the content was modified.
""" """
triggers = list(TRIGGER_PATTERN.finditer(content)) has_inline = bool(INLINE_PATTERN.search(content))
if not triggers: has_document = bool(DOCUMENT_PATTERN.search(content))
if not has_inline and not has_document:
return content, False return content, False
# Process triggers one by one. After each replacement the string length
# may change, so we re-search on the updated content each iteration.
changed = False changed = False
for _ in range(len(triggers)):
match = TRIGGER_PATTERN.search(content) # --- Pass 1: inline triggers ---
# Re-search after each replacement since string length may change.
inline_count = len(INLINE_PATTERN.findall(content))
for _ in range(inline_count):
match = INLINE_PATTERN.search(content)
if not match: if not match:
break break
instruction = match.group(1).strip() instruction = match.group(1).strip()
trigger_line = match.group(0) logger.info("Processing inline trigger: %s", instruction)
logger.info("Processing trigger: %s", instruction)
replacement = _call_with_retry(content, instruction) replacement = _call_with_retry(content, instruction, document_scope=False)
content = content[:match.start()] + replacement + content[match.end():] content = content[:match.start()] + replacement + content[match.end():]
changed = True changed = True
# --- Pass 2: document-scope triggers ---
# Each trigger operates on the result of the previous.
doc_count = len(DOCUMENT_PATTERN.findall(content))
for _ in range(doc_count):
match = DOCUMENT_PATTERN.search(content)
if not match:
break
instruction = match.group(1).strip()
logger.info("Processing document-scope trigger: %s", instruction)
# Remove the trigger line before passing to AI so it doesn't appear
# in the rewritten document. Also consume the trailing newline that
# follows the trigger line, if present.
trigger_start, trigger_end = match.start(), match.end()
if trigger_end < len(content) and content[trigger_end] == "\n":
trigger_end += 1
content_without_trigger = content[:trigger_start] + content[trigger_end:]
result = _call_with_retry(content_without_trigger, instruction, document_scope=True)
if result.startswith("<!-- BLIGHT_FAILED:"):
# On failure, restore the trigger line and insert the failure comment.
content = content[:trigger_start] + result + content[trigger_end:]
else:
content = result
changed = True
return content, changed return content, changed
def _call_with_retry(document: str, instruction: str) -> str: def _call_with_retry(document: str, instruction: str, *, document_scope: bool) -> str:
"""Call the AI provider with up to _MAX_RETRIES attempts. """Call the AI provider with up to _MAX_RETRIES attempts.
Returns the AI response on success, or a BLIGHT_FAILED comment on Returns the AI response on success, or BLIGHT_FAILED/BLIGHT_ERROR comments
exhausted retries. on exhausted retries.
""" """
last_error: Exception | None = None last_error: Exception | None = None
for attempt in range(_MAX_RETRIES): for attempt in range(_MAX_RETRIES):
try: try:
if document_scope:
return _provider.complete_document(document, instruction)
return _provider.complete(document, instruction) return _provider.complete(document, instruction)
except Exception as exc: except Exception as exc:
last_error = exc last_error = exc
@@ -74,4 +116,4 @@ def _call_with_retry(document: str, instruction: str) -> str:
instruction, instruction,
last_error, last_error,
) )
return FAILED_TEMPLATE.format(instruction=instruction) return FAILED_TEMPLATE.format(instruction=instruction, error=last_error)