nniback / core /tools /sb_file_reader_tool.py
Kyrlos11's picture
Upload 842 files
22087d0 verified
Raw History Blame Contribute Delete
26.9 kB
import os
import shlex
import json
import asyncio
from typing import Optional, List
from core.agentpress.tool import ToolResult, openapi_schema, tool_metadata
from core.sandbox.tool_base import SandboxToolsBase
from core.agentpress.thread_manager import ThreadManager
from core.utils.config import config
from core.utils.logger import logger
MAX_OUTPUT_CHARS = 50000
MAX_BATCH_SIZE = 20
KB_VERSION = "0.1.2"
@tool_metadata(
display_name="Read",
description="Read files from the workspace. Access any file directly including code, documents, PDFs, and images",
icon="FileText",
color="bg-emerald-100 dark:bg-emerald-800/50",
is_core=True,
weight=35,
visible=True,
usage_guide="""
## Read - Read files from the workspace
Reads files from the workspace. You can access any file directly by using this tool.
### Key Capabilities
- Reads any file using absolute or relative paths to /workspace
- By default reads up to 2000 lines starting from the beginning
- Supports optional line offset and limit for long files
- Lines longer than 2000 characters will be truncated
- Results are returned with line numbers starting at 1
### File Type Support
- **Images** (PNG, JPG, etc): Contents are presented visually
- **PDF files**: Processed page by page, extracting text and visual content
- **Jupyter notebooks** (.ipynb): Returns all cells with outputs
- **Documents**: PDF, Word (.doc, .docx), PowerPoint (.ppt, .pptx)
- **Data**: Excel (.xls, .xlsx), CSV, JSON, XML
- **Code**: py, js, ts, java, c, cpp, go, rs, etc.
- **Text**: txt, md, log, yaml, toml, ini
### When to Use read_file
- Reading specific files when you know the path
- Small to medium files where you need full content
- Code files where you need complete context
- Config files, scripts, source code
### When to Use search_file
- Large documents where you need specific information
- PDFs, Word docs, Excel files with lots of content
- When searching for specific terms or concepts
### Usage Notes
- The file_path parameter should be relative to /workspace
- Can read multiple files in parallel within a single response
- Use this tool when provided screenshot paths - it works with temporary file paths
- Returns system warning for empty files that exist
### Batch Mode
Both tools support reading/searching multiple files concurrently:
```
read_file(file_paths=["src/main.py", "src/utils.py"])
search_file(file_paths=["docs/report.pdf", "docs/summary.pdf"], query="findings")
```
### Important Notes
- NEVER use cat/head/tail via Bash to read files - use this tool instead
- Multiple files can be read in parallel for efficiency
- For images (jpg, png, gif, webp, svg) → use load_image for visual analysis
"""
)
class SandboxFileReaderTool(SandboxToolsBase):
def __init__(self, project_id: str, thread_manager: ThreadManager):
super().__init__(project_id, thread_manager)
self._kb_ready = False
async def _wait_for_uploads(self, timeout: int = 60) -> bool:
from core.services import redis
key = f"file_upload_pending:{self.project_id}"
start_time = asyncio.get_event_loop().time()
# Reduced polling frequency from 0.5s to 2s to minimize Redis load
POLL_INTERVAL = 2.0
while True:
# Use timeout-protected get
pending = await redis.get(key, timeout=2.0)
if pending is None:
return True
elapsed = asyncio.get_event_loop().time() - start_time
if elapsed >= timeout:
logger.warning(f"[FileReader] Timeout waiting for uploads to complete for project {self.project_id}")
return False
logger.debug(f"[FileReader] Waiting for {pending} file uploads to complete...")
await asyncio.sleep(POLL_INTERVAL)
async def _ensure_kb(self) -> bool:
if self._kb_ready:
return True
try:
check = await self.sandbox.process.exec("kb -v 2>&1")
if check.exit_code == 0 and f"kb-fusion {KB_VERSION}" in check.result:
self._kb_ready = True
return True
url = f"https://github.com/kortix-ai/kb-fusion/releases/download/v{KB_VERSION}/kb"
result = await self.sandbox.process.exec(
f"curl -L -f {url} -o /tmp/kb && chmod +x /tmp/kb && mv /tmp/kb /usr/local/bin/kb"
)
self._kb_ready = result.exit_code == 0
return self._kb_ready
except Exception as e:
logger.error(f"[SearchFile] kb install failed: {e}")
return False
def _get_file_type(self, file_path: str) -> str:
ext = os.path.splitext(file_path)[1].lower()
if ext == '.pdf':
return 'pdf'
elif ext == '.doc':
return 'doc'
elif ext == '.docx':
return 'docx'
elif ext == '.pptx':
return 'pptx'
elif ext == '.ppt':
return 'ppt'
elif ext == '.xlsx':
return 'xlsx'
elif ext == '.xls':
return 'xls'
elif ext in ['.txt', '.md', '.csv', '.json', '.xml', '.html', '.css',
'.js', '.ts', '.py', '.java', '.c', '.cpp', '.h', '.hpp',
'.rb', '.go', '.rs', '.swift', '.kt', '.scala', '.php',
'.sh', '.bash', '.zsh', '.yaml', '.yml', '.toml', '.ini',
'.cfg', '.conf', '.log', '.sql', '.r', '.m', '.tex']:
return 'text'
return 'unknown'
def _is_image_file(self, file_path: str) -> bool:
ext = os.path.splitext(file_path)[1].lower()
return ext in ['.jpg', '.jpeg', '.png', '.gif', '.webp', '.svg', '.bmp', '.ico']
async def _read_single_file(self, file_path: str) -> dict:
try:
cleaned_path = self.clean_path(file_path)
full_path = f"{self.workspace_path}/{cleaned_path}"
if self._is_image_file(cleaned_path):
return {
"file_path": cleaned_path,
"success": False,
"error": f"'{cleaned_path}' is an image file. Use load_image instead of read_file for images."
}
try:
file_info = await self.sandbox.fs.get_file_info(full_path)
if file_info.is_dir:
return {
"file_path": cleaned_path,
"success": False,
"error": f"Path '{cleaned_path}' is a directory, not a file."
}
except Exception:
return {
"file_path": cleaned_path,
"success": False,
"error": f"File not found: '{cleaned_path}'"
}
file_type = self._get_file_type(cleaned_path)
escaped_path = shlex.quote(full_path)
content = ""
extraction_method = "unknown"
if file_type == 'pdf':
# Try pdftotext first (fast for native PDFs)
result = await self.sandbox.process.exec(
f'pdftotext {escaped_path} - 2>/dev/null',
timeout=60
)
if result.exit_code == 0 and result.result.strip():
content = result.result
extraction_method = "pdftotext"
else:
# Try pdftotext with layout preservation
result = await self.sandbox.process.exec(
f'pdftotext -layout {escaped_path} - 2>/dev/null',
timeout=60
)
if result.exit_code == 0 and result.result.strip():
content = result.result
extraction_method = "pdftotext_layout"
else:
# Try PyPDF2 as fallback
result = await self.sandbox.process.exec(
f'python3 -c "import PyPDF2; r=PyPDF2.PdfReader({escaped_path!r}); print(\\"\\n\\".join(p.extract_text() or \\"\\" for p in r.pages))"',
timeout=60
)
if result.exit_code == 0 and result.result.strip():
content = result.result
extraction_method = "PyPDF2"
else:
# Standard extraction failed - likely a scanned/image-only PDF
# Return CLI instructions for manual extraction
return {
"file_path": cleaned_path,
"success": False,
"error": (
f"This PDF ({os.path.basename(cleaned_path)}) is a scanned/image-only document. "
f"Use CLI to extract:\n\n"
f"pdftoppm -png -r 300 '{cleaned_path}' /tmp/page\n"
f"tesseract /tmp/page-1.png stdout -l eng"
)
}
elif file_type == 'doc':
result = await self.sandbox.process.exec(
f'catdoc {escaped_path} 2>/dev/null || antiword {escaped_path} 2>/dev/null',
timeout=60
)
if result.exit_code != 0:
return {
"file_path": cleaned_path,
"success": False,
"error": f"Failed to extract Word doc text"
}
content = result.result
extraction_method = "catdoc"
elif file_type == 'docx':
result = await self.sandbox.process.exec(
f"unzip -p {escaped_path} word/document.xml 2>/dev/null | sed -e 's/<[^>]*>//g' | tr -s ' \\n'",
timeout=60
)
if result.exit_code != 0 or not result.result.strip():
result = await self.sandbox.process.exec(
f'cat {escaped_path}',
timeout=60
)
extraction_method = "raw"
else:
extraction_method = "unzip"
content = result.result
elif file_type == 'pptx':
result = await self.sandbox.process.exec(
f"unzip -p {escaped_path} 'ppt/slides/*.xml' 2>/dev/null | sed -e 's/<[^>]*>//g' | tr -s ' \\n' | head -c 100000",
timeout=60
)
if result.exit_code != 0 or not result.result.strip():
result = await self.sandbox.process.exec(
f"python3 -c \"from pptx import Presentation; p=Presentation({escaped_path!r}); print('\\n'.join(shape.text for slide in p.slides for shape in slide.shapes if hasattr(shape, 'text')))\"",
timeout=60
)
extraction_method = "python-pptx"
else:
extraction_method = "unzip"
if result.exit_code != 0:
return {
"file_path": cleaned_path,
"success": False,
"error": "Failed to extract PowerPoint text. Try installing python-pptx."
}
content = result.result
elif file_type == 'ppt':
result = await self.sandbox.process.exec(
f"catppt {escaped_path} 2>/dev/null",
timeout=60
)
if result.exit_code != 0:
return {
"file_path": cleaned_path,
"success": False,
"error": "Failed to extract old PowerPoint (.ppt). Convert to .pptx for better support."
}
content = result.result
extraction_method = "catppt"
elif file_type == 'xlsx':
result = await self.sandbox.process.exec(
f"python3 -c \"import openpyxl; wb=openpyxl.load_workbook({escaped_path!r}, data_only=True); [print(f'=== Sheet: {{ws.title}} ===') or [print('\\t'.join(str(c.value or '') for c in row)) for row in ws.iter_rows()] for ws in wb.worksheets]\" 2>/dev/null | head -c 100000",
timeout=60
)
if result.exit_code != 0 or not result.result.strip():
result = await self.sandbox.process.exec(
f"unzip -p {escaped_path} 'xl/sharedStrings.xml' 2>/dev/null | sed -e 's/<[^>]*>//g' | tr -s ' \\n' | head -c 100000",
timeout=60
)
extraction_method = "unzip"
else:
extraction_method = "openpyxl"
if result.exit_code != 0:
return {
"file_path": cleaned_path,
"success": False,
"error": "Failed to extract Excel text. Try installing openpyxl."
}
content = result.result
elif file_type == 'xls':
result = await self.sandbox.process.exec(
f"python3 -c \"import xlrd; wb=xlrd.open_workbook({escaped_path!r}); [print(f'=== Sheet: {{ws.name}} ===') or [print('\\t'.join(str(ws.cell_value(r,c)) for c in range(ws.ncols))) for r in range(ws.nrows)] for ws in wb.sheets()]\" 2>/dev/null | head -c 100000",
timeout=60
)
if result.exit_code != 0:
result = await self.sandbox.process.exec(
f"xls2csv {escaped_path} 2>/dev/null | head -c 100000",
timeout=60
)
extraction_method = "xls2csv"
else:
extraction_method = "xlrd"
if result.exit_code != 0:
return {
"file_path": cleaned_path,
"success": False,
"error": "Failed to extract old Excel (.xls). Convert to .xlsx for better support."
}
content = result.result
else:
result = await self.sandbox.process.exec(
f'cat {escaped_path}',
timeout=60
)
if result.exit_code != 0:
return {
"file_path": cleaned_path,
"success": False,
"error": f"Failed to read file"
}
content = result.result
extraction_method = "cat"
truncated = False
if len(content) > MAX_OUTPUT_CHARS:
content = content[:MAX_OUTPUT_CHARS]
truncated = True
content += f"\n\n[Content truncated at {MAX_OUTPUT_CHARS} characters]"
logger.info(f"[ReadFile] Read '{cleaned_path}' ({file_type}) via {extraction_method}, {len(content)} chars")
return {
"file_path": cleaned_path,
"success": True,
"file_type": file_type,
"extraction_method": extraction_method,
"size_bytes": file_info.size,
"content_length": len(content),
"truncated": truncated,
"content": content
}
except Exception as e:
logger.error(f"[ReadFile] Error reading {file_path}: {e}", exc_info=True)
return {
"file_path": file_path,
"success": False,
"error": str(e)
}
@openapi_schema({
"type": "function",
"function": {
"name": "read_file",
"description": """Reads a file from the workspace. You can access any file directly by using this tool.
Usage:
- The file_path parameter should be a path relative to /workspace (e.g., 'src/main.py')
- By default, it reads up to 2000 lines starting from the beginning of the file
- You can optionally specify a line offset and limit for long files
- Any lines longer than 2000 characters will be truncated
- Results are returned with line numbers starting at 1
- This tool can read images (PNG, JPG), PDFs, Jupyter notebooks (.ipynb), and all text-based files
- Can read multiple files in parallel for efficiency""",
"parameters": {
"type": "object",
"properties": {
"file_path": {
"type": "string",
"description": "**OPTIONAL** - Single file path within /workspace. Example: 'src/main.py'. Use this OR file_paths, not both."
},
"file_paths": {
"type": "array",
"items": {"type": "string"},
"description": "**OPTIONAL** - Array of file paths to read concurrently. Example: ['src/main.py', 'src/utils.py']. Max 20 files per batch. Use this OR file_path, not both."
},
"offset": {
"type": "integer",
"description": "**OPTIONAL** - The line number to start reading from (1-indexed). Only provide if the file is too large to read at once."
},
"limit": {
"type": "integer",
"description": "**OPTIONAL** - The number of lines to read. Only provide if the file is too large to read at once."
}
},
"additionalProperties": False
}
}
})
async def read_file(
self,
file_path: Optional[str] = None,
file_paths: Optional[List[str]] = None,
offset: Optional[int] = None,
limit: Optional[int] = None
) -> ToolResult:
try:
await self._ensure_sandbox()
await self._wait_for_uploads()
if file_paths:
paths_to_read = file_paths[:MAX_BATCH_SIZE]
if len(file_paths) > MAX_BATCH_SIZE:
logger.warning(f"[ReadFile] Batch limited to {MAX_BATCH_SIZE} files, {len(file_paths)} provided")
elif file_path:
paths_to_read = [file_path]
else:
return self.fail_response("Either 'file_path' or 'file_paths' must be provided.")
if len(paths_to_read) == 1:
result = await self._read_single_file(paths_to_read[0])
if result.get("success"):
return self.success_response(result)
else:
return self.fail_response(result.get("error", "Unknown error"))
tasks = [self._read_single_file(path) for path in paths_to_read]
results = await asyncio.gather(*tasks)
successful = [r for r in results if r.get("success")]
failed = [r for r in results if not r.get("success")]
logger.info(f"[ReadFile] Batch read: {len(successful)} success, {len(failed)} failed out of {len(paths_to_read)} files")
return self.success_response({
"batch_mode": True,
"total_files": len(paths_to_read),
"successful": len(successful),
"failed": len(failed),
"results": results
})
except Exception as e:
logger.error(f"[ReadFile] Error: {e}", exc_info=True)
return self.fail_response(f"Error reading files: {str(e)}")
@openapi_schema({
"type": "function",
"function": {
"name": "search_file",
"description": """A powerful semantic search tool for finding content within files.
Usage:
- ALWAYS use search_file for large documents. NEVER use grep or rg as a Bash command.
- This tool has been optimized for correct permissions and access.
- Returns relevant chunks instead of full content - prevents context flooding.
- Supports semantic search with natural language queries.
When to use:
- Large PDFs and documents where you need specific information
- Finding specific terms, concepts, or patterns in files
- Multiple files where you need to search for something
- When read_file would return too much content
Examples:
- search_file(file_path="docs/contract.pdf", query="termination clause")
- search_file(file_paths=["doc1.pdf", "doc2.pdf"], query="payment terms")""",
"parameters": {
"type": "object",
"properties": {
"query": {
"type": "string",
"description": "**REQUIRED** - The search query. Can be natural language ('What are the payment terms?') or specific terms ('termination clause')."
},
"file_path": {
"type": "string",
"description": "**OPTIONAL** - Single file path to search within. Example: 'docs/contract.pdf'"
},
"file_paths": {
"type": "array",
"items": {"type": "string"},
"description": "**OPTIONAL** - Multiple file paths to search across. Example: ['docs/report1.pdf', 'docs/report2.pdf']"
}
},
"required": ["query"],
"additionalProperties": False
}
}
})
async def search_file(
self,
query: str,
file_path: Optional[str] = None,
file_paths: Optional[List[str]] = None
) -> ToolResult:
try:
await self._ensure_sandbox()
await self._wait_for_uploads()
if not query or not query.strip():
return self.fail_response("Query is required for search.")
paths = []
if file_paths:
paths = [f"{self.workspace_path}/{self.clean_path(p)}" for p in file_paths]
elif file_path:
paths = [f"{self.workspace_path}/{self.clean_path(file_path)}"]
else:
# Check default paths: uploads, KB directory, and workspace root
default_paths = [
f"{self.workspace_path}/uploads",
f"{self.workspace_path}/downloads/global-knowledge",
self.workspace_path
]
for candidate in default_paths:
check = await self.sandbox.process.exec(f"test -e {shlex.quote(candidate)} && echo 'ok'")
if "ok" in check.result:
paths.append(candidate)
if not paths:
return self.fail_response(
"No searchable paths found. Upload files to /workspace/uploads or sync knowledge base with global_kb_sync first."
)
# Verify explicitly provided paths exist
if file_paths or file_path:
for p in paths:
check = await self.sandbox.process.exec(f"test -e {shlex.quote(p)} && echo 'ok'")
if "ok" not in check.result:
return self.fail_response(f"Path not found: {p}")
if not await self._ensure_kb():
return self.fail_response("Failed to initialize search. Try read_file instead.")
env = {"OPENAI_API_KEY": config.OPENAI_API_KEY} if config.OPENAI_API_KEY else {}
path_args = " ".join([shlex.quote(p) for p in paths])
escaped_query = shlex.quote(query)
cmd = f"kb search {path_args} {escaped_query} -k 10 --json"
result = await self.sandbox.process.exec(cmd, env=env, cwd=self.workspace_path, timeout=120)
if result.exit_code != 0:
logger.error(f"[SearchFile] kb search failed: {result.result}")
return self.fail_response(f"Search failed: {result.result[:500]}")
logger.info(f"[SearchFile] Raw kb output (first 1000 chars): {result.result[:1000]}")
try:
search_results = json.loads(result.result)
total_hits = 0
formatted_results = []
for query_result in search_results:
q = query_result.get("query", query)
hits = query_result.get("hits", [])
total_hits += len(hits)
if hits:
logger.info(f"[SearchFile] First hit structure: {list(hits[0].keys())}")
for hit in hits[:10]:
content = hit.get("content") or hit.get("chunk") or hit.get("text") or hit.get("snippet") or ""
file_path = hit.get("file_path") or hit.get("path") or hit.get("file") or ""
if not content and hit:
logger.info(f"[SearchFile] Hit has no content. Keys: {list(hit.keys())}, first 500 chars: {str(hit)[:500]}")
formatted_results.append({
"file": file_path.replace(self.workspace_path + "/", ""),
"score": round(hit.get("score", 0), 3),
"content": content[:8000] if content else "[No content - check hit structure]"
})
logger.info(f"[SearchFile] Found {total_hits} results for '{query}'")
return self.success_response({
"query": query,
"total_hits": total_hits,
"results": formatted_results,
"note": "Showing top relevant chunks. Use read_file if you need the complete file."
})
except json.JSONDecodeError:
return self.success_response({
"query": query,
"raw_results": result.result[:5000]
})
except Exception as e:
logger.error(f"[SearchFile] Error: {e}", exc_info=True)
return self.fail_response(f"Search error: {str(e)}")