alita-sdk 0.3.465__py3-none-any.whl → 0.3.497__py3-none-any.whl

alita_sdk/tools/chunkers/sematic/markdown_chunker.py

@@ -1,4 +1,4 @@
-from typing import Generator
+from typing import Generator, List
 from langchain_core.documents import Document
 from langchain_text_splitters import MarkdownHeaderTextSplitter, ExperimentalMarkdownSyntaxTextSplitter
 from langchain.text_splitter import TokenTextSplitter
@@ -7,34 +7,60 @@ from copy import deepcopy as copy
 
 
 def markdown_chunker(file_content_generator: Generator[Document, None, None], config: dict, *args, **kwargs) -> Generator[Document, None, None]:
+    """
+    Chunks markdown documents by headers, with support for:
+    - Minimum chunk size to avoid tiny fragments
+    - Maximum token limit with overflow splitting
+    - Header metadata preservation
+    
+    Config options:
+        strip_header (bool): Remove headers from content. Default: False
+        return_each_line (bool): Split on every line. Default: False
+        headers_to_split_on (list): Headers to split on, e.g. [('#', 'H1'), ('##', 'H2')]
+        max_tokens (int): Maximum tokens per chunk. Default: 512
+        token_overlap (int): Token overlap for large chunk splitting. Default: 10
+        min_chunk_chars (int): Minimum characters per chunk. Default: 100
+            Chunks smaller than this will be merged with the next chunk.
+    """
     strip_header = config.get("strip_header", False)
     return_each_line = config.get("return_each_line", False)
     headers_to_split_on = config.get("headers_to_split_on", [])
     max_tokens = config.get("max_tokens", 512)
     tokens_overlapping = config.get("token_overlap", 10)
+    min_chunk_chars = config.get("min_chunk_chars", 100)  # Minimum characters per chunk
+    
     headers_to_split_on = [tuple(header) for header in headers_to_split_on]
+    
     for doc in file_content_generator:
         doc_metadata = doc.metadata
         doc_content = doc.page_content
         chunk_id = 0
+        
         markdown_splitter = MarkdownHeaderTextSplitter(
             headers_to_split_on=headers_to_split_on, 
             strip_headers=strip_header,
             return_each_line=return_each_line
         )
         md_header_splits = markdown_splitter.split_text(doc_content)
-        for chunk in md_header_splits:
+        
+        # Merge small chunks with the next one
+        merged_chunks = _merge_small_chunks(md_header_splits, min_chunk_chars)
+        
+        for chunk in merged_chunks:
             if tiktoken_length(chunk.page_content) > max_tokens:
-                for subchunk in TokenTextSplitter(encoding_name="cl100k_base", 
-                                                  chunk_size=max_tokens, 
-                                                  chunk_overlap=tokens_overlapping
-                                                  ).split_text(chunk.page_content):
+                # Split large chunks into smaller ones
+                for subchunk in TokenTextSplitter(
+                    encoding_name="cl100k_base", 
+                    chunk_size=max_tokens, 
+                    chunk_overlap=tokens_overlapping
+                ).split_text(chunk.page_content):
                     chunk_id += 1
                     headers_meta = list(chunk.metadata.values())
                     docmeta = copy(doc_metadata)
                     docmeta.update({"headers": "; ".join(headers_meta)})
                     docmeta['chunk_id'] = chunk_id
                     docmeta['chunk_type'] = "document"
+                    docmeta['method_name'] = 'markdown'
                     yield Document(
                         page_content=subchunk,
                         metadata=docmeta
@@ -46,12 +72,77 @@ def markdown_chunker(file_content_generator: Generator[Document, None, None], co
                 docmeta.update({"headers": "; ".join(headers_meta)})
                 docmeta['chunk_id'] = chunk_id
                 docmeta['chunk_type'] = "document"
+                docmeta['method_name'] = 'text'
                 yield Document(
                     page_content=chunk.page_content,
                     metadata=docmeta
                 )
 
 
+def _merge_small_chunks(chunks: List[Document], min_chars: int) -> List[Document]:
+    """
+    Merge chunks that are smaller than min_chars with the next chunk.
+    
+    This prevents tiny fragments (like standalone headers or short notes)
+    from becoming separate chunks.
+    
+    Args:
+        chunks: List of Document chunks from markdown splitter
+        min_chars: Minimum character count for a chunk
+        
+    Returns:
+        List of merged Document chunks
+    """
+    if not chunks:
+        return chunks
+    
+    merged = []
+    pending_content = ""
+    pending_metadata = {}
+    
+    for i, chunk in enumerate(chunks):
+        content = chunk.page_content.strip()
+        
+        if pending_content:
+            # Merge pending content with current chunk
+            combined_content = pending_content + "\n\n" + content
+            # Use the pending metadata (from the header) but can be extended
+            combined_metadata = {**pending_metadata}
+            # Add any new header info from current chunk
+            for key, value in chunk.metadata.items():
+                if key not in combined_metadata or not combined_metadata[key]:
+                    combined_metadata[key] = value
+            
+            if len(combined_content) >= min_chars:
+                # Combined is big enough, emit it
+                merged.append(Document(
+                    page_content=combined_content,
+                    metadata=combined_metadata
+                ))
+                pending_content = ""
+                pending_metadata = {}
+            else:
+                # Still too small, keep accumulating
+                pending_content = combined_content
+                pending_metadata = combined_metadata
+        elif len(content) < min_chars:
+            # Current chunk is too small, start pending
+            pending_content = content
+            pending_metadata = dict(chunk.metadata)
+        else:
+            # Current chunk is big enough
+            merged.append(chunk)
+    
+    # Don't forget any remaining pending content
+    if pending_content:
+        merged.append(Document(
+            page_content=pending_content,
+            metadata=pending_metadata
+        ))
+    
+    return merged
+
+
 def markdown_by_headers_chunker(file_content_generator: Generator[Document, None, None], config: dict, *args, **kwargs) -> Generator[Document, None, None]:
     strip_header = config.get("strip_header", False)
     return_each_line = config.get("return_each_line", False)

alita_sdk/tools/chunkers/universal_chunker.py

@@ -0,0 +1,270 @@
+"""
+Universal Chunker - Routes documents to appropriate chunkers based on file type.
+
+This module provides a universal chunking interface that automatically selects
+the appropriate chunking strategy based on the file extension:
+
+- .md, .markdown → Markdown chunker (header-based splitting)
+- .py, .js, .ts, .java, etc. → TreeSitter code chunker
+- .json → JSON chunker  
+- other → Default text chunker
+
+Usage:
+    from alita_sdk.tools.chunkers.universal_chunker import universal_chunker
+    
+    # Chunk documents from a loader
+    for chunk in universal_chunker(document_generator, config):
+        print(chunk.page_content)
+"""
+
+import logging
+import os
+from typing import Generator, Dict, Any, Optional
+from langchain_core.documents import Document
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+
+from .code.codeparser import parse_code_files_for_db
+from .sematic.markdown_chunker import markdown_chunker
+from .sematic.json_chunker import json_chunker
+
+logger = logging.getLogger(__name__)
+
+
+# File extension mappings
+MARKDOWN_EXTENSIONS = {'.md', '.markdown', '.mdown', '.mkd', '.mdx'}
+JSON_EXTENSIONS = {'.json', '.jsonl', '.jsonc'}
+CODE_EXTENSIONS = {
+    '.py', '.js', '.jsx', '.mjs', '.cjs', '.ts', '.tsx',
+    '.java', '.kt', '.rs', '.go', '.cpp', '.c', '.cs', 
+    '.hs', '.rb', '.scala', '.lua'
+}
+
+
+def get_file_extension(file_path: str) -> str:
+    """Extract file extension from path."""
+    return os.path.splitext(file_path)[-1].lower()
+
+
+def get_file_type(file_path: str) -> str:
+    """
+    Determine the file type category for chunking.
+    
+    Returns:
+        'markdown', 'json', 'code', or 'text'
+    """
+    ext = get_file_extension(file_path)
+    
+    if ext in MARKDOWN_EXTENSIONS:
+        return 'markdown'
+    elif ext in JSON_EXTENSIONS:
+        return 'json'
+    elif ext in CODE_EXTENSIONS:
+        return 'code'
+    else:
+        return 'text'
+
+
+def _default_text_chunker(
+    documents: Generator[Document, None, None], 
+    config: Dict[str, Any]
+) -> Generator[Document, None, None]:
+    """
+    Default text chunker for unknown file types.
+    Uses recursive character splitting.
+    """
+    chunk_size = config.get('chunk_size', 1000)
+    chunk_overlap = config.get('chunk_overlap', 100)
+    
+    splitter = RecursiveCharacterTextSplitter(
+        chunk_size=chunk_size,
+        chunk_overlap=chunk_overlap,
+        length_function=len,
+    )
+    
+    for doc in documents:
+        chunks = splitter.split_documents([doc])
+        for idx, chunk in enumerate(chunks, 1):
+            chunk.metadata['chunk_id'] = idx
+            chunk.metadata['chunk_type'] = 'text'
+            chunk.metadata['method_name'] = 'text'
+            yield chunk
+
+
+def _code_chunker_from_documents(
+    documents: Generator[Document, None, None],
+    config: Dict[str, Any]
+) -> Generator[Document, None, None]:
+    """
+    Adapter to convert Document generator to code parser format.
+    """
+    def file_content_generator():
+        for doc in documents:
+            yield {
+                'file_name': doc.metadata.get('file_path', doc.metadata.get('filename', 'unknown')),
+                'file_content': doc.page_content,
+                'commit_hash': doc.metadata.get('commit_hash', ''),
+            }
+    
+    # parse_code_files_for_db returns chunks with proper metadata
+    for chunk in parse_code_files_for_db(file_content_generator()):
+        # Ensure file_path is preserved
+        if 'file_path' not in chunk.metadata and 'filename' in chunk.metadata:
+            chunk.metadata['file_path'] = chunk.metadata['filename']
+        yield chunk
+
+
+def universal_chunker(
+    documents: Generator[Document, None, None],
+    config: Optional[Dict[str, Any]] = None
+) -> Generator[Document, None, None]:
+    """
+    Universal chunker that routes documents to appropriate chunkers based on file type.
+    
+    Each document is inspected for its file extension (from metadata.file_path or
+    metadata.file_name) and routed to the appropriate chunker:
+    
+    - Markdown files → markdown_chunker (header-based splitting)
+    - JSON files → json_chunker (recursive JSON splitting)
+    - Code files → code parser (TreeSitter-based parsing)
+    - Other files → default text chunker (recursive character splitting)
+    
+    Args:
+        documents: Generator yielding Document objects with file content
+        config: Optional configuration dict with:
+            - markdown_config: Config for markdown chunker
+            - json_config: Config for JSON chunker
+            - code_config: Config for code chunker
+            - text_config: Config for default text chunker
+            
+    Yields:
+        Document objects with chunked content and preserved metadata
+    """
+    if config is None:
+        config = {}
+    
+    # Default configs for each chunker type
+    markdown_config = config.get('markdown_config', {
+        'strip_header': False,
+        'return_each_line': False,
+        'headers_to_split_on': [
+            ('#', 'Header 1'),
+            ('##', 'Header 2'),
+            ('###', 'Header 3'),
+            ('####', 'Header 4'),
+        ],
+        'max_tokens': 1024,
+        'token_overlap': 50,
+        'min_chunk_chars': 100,  # Merge chunks smaller than this
+    })
+    
+    json_config = config.get('json_config', {
+        'max_tokens': 512,
+    })
+    
+    code_config = config.get('code_config', {})
+    
+    text_config = config.get('text_config', {
+        'chunk_size': 1000,
+        'chunk_overlap': 100,
+    })
+    
+    # Buffer documents by type for batch processing
+    # This is more efficient than processing one at a time
+    markdown_docs = []
+    json_docs = []
+    code_docs = []
+    text_docs = []
+    
+    # Buffer size before flushing
+    BUFFER_SIZE = 10
+    
+    def flush_markdown():
+        if markdown_docs:
+            def gen():
+                for d in markdown_docs:
+                    yield d
+            for chunk in markdown_chunker(gen(), markdown_config):
+                yield chunk
+            markdown_docs.clear()
+    
+    def flush_json():
+        if json_docs:
+            def gen():
+                for d in json_docs:
+                    yield d
+            for chunk in json_chunker(gen(), json_config):
+                yield chunk
+            json_docs.clear()
+    
+    def flush_code():
+        if code_docs:
+            def gen():
+                for d in code_docs:
+                    yield d
+            for chunk in _code_chunker_from_documents(gen(), code_config):
+                yield chunk
+            code_docs.clear()
+    
+    def flush_text():
+        if text_docs:
+            def gen():
+                for d in text_docs:
+                    yield d
+            for chunk in _default_text_chunker(gen(), text_config):
+                yield chunk
+            text_docs.clear()
+    
+    for doc in documents:
+        # Get file path from metadata
+        file_path = (doc.metadata.get('file_path') or 
+                    doc.metadata.get('file_name') or 
+                    doc.metadata.get('source') or 
+                    'unknown')
+        
+        # Ensure file_path is in metadata for downstream use
+        doc.metadata['file_path'] = file_path
+        
+        file_type = get_file_type(file_path)
+        
+        if file_type == 'markdown':
+            markdown_docs.append(doc)
+            if len(markdown_docs) >= BUFFER_SIZE:
+                yield from flush_markdown()
+        elif file_type == 'json':
+            json_docs.append(doc)
+            if len(json_docs) >= BUFFER_SIZE:
+                yield from flush_json()
+        elif file_type == 'code':
+            code_docs.append(doc)
+            if len(code_docs) >= BUFFER_SIZE:
+                yield from flush_code()
+        else:
+            text_docs.append(doc)
+            if len(text_docs) >= BUFFER_SIZE:
+                yield from flush_text()
+    
+    # Flush remaining documents
+    yield from flush_markdown()
+    yield from flush_json()
+    yield from flush_code()
+    yield from flush_text()
+
+
+def chunk_single_document(
+    doc: Document,
+    config: Optional[Dict[str, Any]] = None
+) -> Generator[Document, None, None]:
+    """
+    Convenience function to chunk a single document.
+    
+    Args:
+        doc: Single Document to chunk
+        config: Optional chunker configuration
+        
+    Yields:
+        Chunked Document objects
+    """
+    def single_doc_gen():
+        yield doc
+    
+    yield from universal_chunker(single_doc_gen(), config)

alita_sdk/tools/code/loaders/codesearcher.py

@@ -4,8 +4,9 @@ def search_format(items):
     results = []
     for (doc, score) in items:
         res_chunk = ''
-        language = get_programming_language(get_file_extension(doc.metadata["filename"]))
-        res_chunk += doc.metadata["filename"] + " -> " + doc.metadata["method_name"] + " (score: " + str(score) + ")"
+        language = get_programming_language(get_file_extension(doc.metadata.get("filename", "unknown")))
+        method_name = doc.metadata.get("method_name", "text")
+        res_chunk += doc.metadata.get("filename", "unknown") + " -> " + method_name + " (score: " + str(score) + ")"
         res_chunk += "\n\n```" + language.value + "\n"+ doc.page_content + "\n```\n\n"
         results.append(res_chunk)
     return results

alita_sdk/tools/code_indexer_toolkit.py

@@ -9,13 +9,13 @@ from langchain_core.tools import ToolException
 from pydantic import Field
 
 from alita_sdk.tools.base_indexer_toolkit import BaseIndexerToolkit
-from .chunkers.code.codeparser import parse_code_files_for_db
 
 logger = logging.getLogger(__name__)
 
 
 class CodeIndexerToolkit(BaseIndexerToolkit):
     def _get_indexed_data(self, index_name: str):
+        self._ensure_vectorstore_initialized()
         if not self.vector_adapter:
             raise ToolException("Vector adapter is not initialized. "
                              "Check your configuration: embedding_model and vectorstore_type.")
@@ -66,26 +66,40 @@ class CodeIndexerToolkit(BaseIndexerToolkit):
     def loader(self,
                branch: Optional[str] = None,
                whitelist: Optional[List[str]] = None,
-               blacklist: Optional[List[str]] = None) -> Generator[Document, None, None]:
+               blacklist: Optional[List[str]] = None,
+               chunked: bool = True) -> Generator[Document, None, None]:
         """
-        Generates file content from a branch, respecting whitelist and blacklist patterns.
+        Generates Documents from files in a branch, respecting whitelist and blacklist patterns.
     
         Parameters:
         - branch (Optional[str]): Branch for listing files. Defaults to the current branch if None.
         - whitelist (Optional[List[str]]): File extensions or paths to include. Defaults to all files if None.
         - blacklist (Optional[List[str]]): File extensions or paths to exclude. Defaults to no exclusions if None.
+        - chunked (bool): If True (default), applies universal chunker based on file type.
+                         If False, returns raw Documents without chunking.
     
         Returns:
-        - generator: Yields content from files matching the whitelist but not the blacklist.
+        - generator: Yields Documents from files matching the whitelist but not the blacklist.
+                    Each document has exactly the key 'filename' in metadata, which is used as an ID
+                    for further operations (indexing, deduplication, and retrieval).
     
         Example:
         # Use 'feature-branch', include '.py' files, exclude 'test_' files
-        file_generator = loader(branch='feature-branch', whitelist=['*.py'], blacklist=['*test_*'])
+        for doc in loader(branch='feature-branch', whitelist=['*.py'], blacklist=['*test_*']):
+            print(doc.page_content)
     
         Notes:
         - Whitelist and blacklist use Unix shell-style wildcards.
         - Files must match the whitelist and not the blacklist to be included.
+        - Each document MUST have exactly the key 'filename' in metadata. This key is used as an ID
+          for further operations such as indexing, deduplication, and retrieval.
+        - When chunked=True:
+          - .md files → markdown chunker (header-based splitting)
+          - .py/.js/.ts/etc → code parser (TreeSitter-based)
+          - .json files → JSON chunker
+          - other files → default text chunker
         """
+        import hashlib
     
         _files = self.__handle_get_files("", self.__get_branch(branch))
         self._log_tool_event(message="Listing files in branch", tool_name="loader")
@@ -103,41 +117,60 @@ class CodeIndexerToolkit(BaseIndexerToolkit):
                         or any(file_path.endswith(f'.{pattern}') for pattern in blacklist))
             return False
     
-        def file_content_generator():
+        def raw_document_generator() -> Generator[Document, None, None]:
+            """Yields raw Documents without chunking."""
             self._log_tool_event(message="Reading the files", tool_name="loader")
-            # log the progress of file reading
             total_files = len(_files)
+            processed = 0
+            
             for idx, file in enumerate(_files, 1):
                 if is_whitelisted(file) and not is_blacklisted(file):
-                    # read file ONLY if it matches whitelist and does not match blacklist
                     try:
                         file_content = self._read_file(file, self.__get_branch(branch))
                     except Exception as e:
                         logger.error(f"Failed to read file {file}: {e}")
-                        file_content = ""
+                        continue
+                    
                     if not file_content:
-                        # empty file, skip
                         continue
-                    #
-                    # ensure file content is a string
+                    
+                    # Ensure file content is a string
                     if isinstance(file_content, bytes):
                         file_content = file_content.decode("utf-8", errors="ignore")
                     elif isinstance(file_content, dict) and file.endswith('.json'):
                         file_content = json.dumps(file_content)
                     elif not isinstance(file_content, str):
                         file_content = str(file_content)
-                    #
-                    # hash the file content to ensure uniqueness
-                    import hashlib
+                    
+                    # Hash the file content for uniqueness tracking
                     file_hash = hashlib.sha256(file_content.encode("utf-8")).hexdigest()
-                    yield {"file_name": file,
-                           "file_content": file_content,
-                           "commit_hash": file_hash}
+                    processed += 1
+                    
+                    yield Document(
+                        page_content=file_content,
+                        metadata={
+                            'file_path': file,
+                            'filename': file,
+                            'source': file,
+                            'commit_hash': file_hash,
+                        }
+                    )
+                
                 if idx % 10 == 0 or idx == total_files:
-                    self._log_tool_event(message=f"{idx} out of {total_files} files have been read", tool_name="loader")
-            self._log_tool_event(message=f"{len(_files)} have been read", tool_name="loader")
-    
-        return parse_code_files_for_db(file_content_generator())
+                    self._log_tool_event(
+                        message=f"{idx} out of {total_files} files checked, {processed} matched",
+                        tool_name="loader"
+                    )
+            
+            self._log_tool_event(message=f"{processed} files loaded", tool_name="loader")
+
+        if not chunked:
+            # Return raw documents without chunking
+            return raw_document_generator()
+        
+        # Apply universal chunker based on file type
+        from .chunkers.universal_chunker import universal_chunker
+        return universal_chunker(raw_document_generator())
 
     def __handle_get_files(self, path: str, branch: str):
         """

alita_sdk/tools/confluence/api_wrapper.py

@@ -480,21 +480,69 @@ class ConfluenceAPIWrapper(NonCodeIndexerToolkit):
         """Gets pages with specific label in the Confluence space."""
 
         start = 0
-        pages_info = []
-        for _ in range((self.max_pages + self.limit - 1) // self.limit):
-            pages = self.client.get_all_pages_by_label(label, start=start,
-                                                       limit=self.limit)  # , expand="body.view.value"
+        pages_info: List[Dict[str, Any]] = []
+        seen_ids: set[str] = set()
+
+        # Use a while-loop driven by unique pages collected and
+        # presence of additional results instead of a fixed number
+        # of iterations based purely on max_pages/limit.
+        while len(pages_info) < (self.max_pages or 0):
+            pages = self.client.get_all_pages_by_label(
+                label,
+                start=start,
+                limit=self.limit,
+            )  # , expand="body.view.value"
             if not pages:
                 break
 
-            pages_info += [{
-                'page_id': page.metadata['id'],
-                'page_title': page.metadata['title'],
-                'page_url': page.metadata['source'],
-                'content': page.page_content
-            } for page in self.get_pages_by_id([page["id"] for page in pages])]
+            # Collect only ids we haven't processed yet to avoid
+            # calling get_page_by_id multiple times for the same
+            # Confluence page.
+            new_ids: List[str] = []
+            for p in pages:
+                page_id = p["id"] if isinstance(p, dict) else getattr(p, "id", None)
+                if page_id is None:
+                    continue
+                if page_id in seen_ids:
+                    continue
+                seen_ids.add(page_id)
+                new_ids.append(page_id)
+
+            if new_ids:
+                for page in self.get_pages_by_id(new_ids):
+                    meta = getattr(page, "metadata", {}) or {}
+                    page_id = meta.get("id")
+                    page_title = meta.get("title")
+                    page_url = meta.get("source")
+                    content = getattr(page, "page_content", None)
+
+                    if page_id is None:
+                        continue
+
+                    pages_info.append(
+                        {
+                            "page_id": page_id,
+                            "page_title": page_title,
+                            "page_url": page_url,
+                            "content": content,
+                        }
+                    )
+
+                    # Respect max_pages on unique pages collected.
+                    if len(pages_info) >= (self.max_pages or 0):
+                        break
+
+            # Advance the offset by the requested page size.
             start += self.limit
-        return pages_info
+
+            # Defensive break: if the API returns fewer items than
+            # requested, there are likely no more pages to fetch.
+            if len(pages) < self.limit:
+                break
+
+        # Slice as an extra safety net in case of any race conditions
+        # around the max_pages guard in the loop above.
+        return pages_info[: (self.max_pages or len(pages_info))]
 
     def is_public_page(self, page: dict) -> bool:
         """Check if a page is publicly accessible."""
@@ -896,14 +944,14 @@ class ConfluenceAPIWrapper(NonCodeIndexerToolkit):
 
                     # Re-verify extension filters
                     # Check if file should be skipped based on skip_extensions
-                    if any(re.match(pattern.replace('*', '.*') + '$', title, re.IGNORECASE)
+                    if any(re.match(re.escape(pattern).replace(r'\*', '.*') + '$', title, re.IGNORECASE)
                            for pattern in self._skip_extensions):
                         continue
 
                     # Check if file should be included based on include_extensions
                     # If include_extensions is empty, process all files (that weren't skipped)
                     if self._include_extensions and not (
-                    any(re.match(pattern.replace('*', '.*') + '$', title, re.IGNORECASE)
+                    any(re.match(re.escape(pattern).replace(r'\*', '.*') + '$', title, re.IGNORECASE)
                         for pattern in self._include_extensions)):
                         continue
 
@@ -1820,4 +1868,5 @@ class ConfluenceAPIWrapper(NonCodeIndexerToolkit):
                 "description": self.get_page_attachments.__doc__,
                 "args_schema": GetPageAttachmentsInput,
             }
-        ]
+        ]
+