ace-context-to-memory（王牌记忆上下文）

Ace-Context-to-Memory

Converts source documents to markdown and chunks them for agent memory. Pipeline: convert (documents → markdown + images) → chunk (markdown → smaller files) → embed (vector index for semantic search).

When to Activate

• User asks to "add content to memory" or "refresh memory"
• Wants to convert a folder of documents (PPTX, PDF, DOCX, XLSX) for agent context
• Mentions "convert and chunk", "ingest for agent", or "memory pipeline"
• Has added new files and wants them processed
• "Use memory" / semantic retrieval: User says "use memory", "search memory", "what does memory say about X", "from our content", "from ABD materials", "what do we have on [topic]" — run search_memory "<query>" and inject results into your response
• Adding content from any folder: When adding a folder's content to memory, run link_workspace_source.py --path <folder> first (on request) if the link does not yet exist

CRITICAL: Respect User Scope

• One file: When user says "one file", "this file", "just X.pdf", or names a specific file → use --file <path>. Process ONLY that file.
• Folder: When user says "folder", "all", "everything in X", or explicitly requests a folder → use --memory <path>.
• Do NOT process entire folders when the user asked for a single file.

OneDrive Folder Clarification

When the user refers to "OneDrive", "OneDrive folder", "slash OneDrive", or similar without specifying which OneDrive:

• Ask which OneDrive folder they mean (e.g. "OneDrive - Agile by Design", "OneDrive - Personal", "OneDrive - Company Name").
• If multiple OneDrive folders exist under the user's home and the configured path is missing or unclear, scripts will list them and prompt for choice — do not guess; ask the user to pick.
• If the user's request is ambiguous (e.g. "use OneDrive" or "get to OneDrive"), ask: "Which OneDrive folder do you want to use? For example: OneDrive - Agile by Design, OneDrive - Personal, or another?"

Source Folder Structure

source/ (under workspace root) contains content sources. Workspace folders can be linked for skill access:

• source/<link_name> — Junction/symlink to your content folder

Use link_workspace_source.py to create links before converting. Set CONTENT_MEMORY_ROOT if workspace root differs from cwd.

Pipeline Overview

1. Convert: Use markitdown to convert supported files to markdown. Images extracted and referenced. SharePoint links: When source is in OneDrive, SharePoint URLs are auto-injected from sharepoint_mapping.json so links work for anyone.
2. Chunk: Split large markdown by slides (decks) or headings (docs). Small files stay as single chunks.
3. Sync SharePoint URLs: Replace source paths with SharePoint URLs; fix URL order; add wdSlideIndex (pptx) / page (pdf) for direct slide/page links. Run automatically in index_memory --path pipeline.
4. Embed + Index (RAG): Embed chunks with sentence-transformers, store in ChromaDB for semantic search.

Semantic Search (RAG)

When the user says "use memory", "search memory", "what does memory say about X", "from our content", or asks about ABD materials:

1. Run from workspace root: python skills/ace-context-to-memory/scripts/search_memory.py "<query>" --k 5
2. Inject the returned chunks into your response
3. Cite sources (path, slide/page) when using retrieved content

Requires RAG deps: pip install -r skills/ace-context-to-memory/requirements-rag.txt

See content/rag-retrieval.md for trigger phrases and agent flow.

Scripts

Run from workspace root. Scripts in skills/ace-context-to-memory/scripts/.

RAG (vector search):

• index_memory.py --path <source_folder> — full pipeline: convert → chunk → sync SharePoint → embed (or chunk + embed if convert already ran)
• index_memory.py --replace — rebuild entire vector index from all memory
• search_memory.py "<query>" [--k 5] [--format text|json] — semantic search; returns top-k chunks

Convert + chunk:

• link_workspace_source.py --path <folder> [--name <link_name>] or --workspace <folder_name> — run on request when adding content to memory; creates junction/symlink in source/ so skills can access the folder
• convert_to_markdown.py --file <file_path> — single file only (use when user asks for one file); writes markdown alongside the source file (same folder)
• convert_to_markdown.py --memory <source_path> — folder (all supported files). When source is in OneDrive, SharePoint URLs are auto-injected via sharepoint_mapping.json.
• chunk_markdown.py --path <source_folder> [--memory <memory_name>] — reads from source, writes to memory//
• embed_and_index.py [--memory <memory_name>] [--replace] — embed chunks from memory// into FAISS index (called by index_memory)
• sync_sharepoint_urls.py [--memory <memory_name>] — run after chunk when source has source/... | https://...; replaces with SharePoint URL, fixes URL order, adds wdSlideIndex (pptx) / page (pdf) for direct links
• add_sharepoint_mapping.py --prefix "OneDrive - X" --base "<url>" — add OneDrive→SharePoint mapping. Paste URL from browser; script derives base. Use when convert warns about missing mapping.

Export (markdown → Excel, Word, PDF):

• markdown_to_excel.py <input.md> [output.xlsx] — generic md → Excel (headings, tables, paragraphs)
• markdown_to_docx.py <input.md> [output.docx] — md → Word (requires pandoc)
• markdown_to_pdf.py <input.md> [output.pdf] — md → PDF (requires pandoc + PDF engine). Use --pdf-engine weasyprint if pdflatex not installed.

Requires: pip install -r requirements-export.txt (openpyxl, pypandoc). Project-specific exports (e.g. JBOM format) stay in workspace.

When adding content from a folder to memory: Run link_workspace_source.py first if the link does not yet exist. Examples:

python scripts/link_workspace_source.py --path "C:/docs/RFQ materials" --name "JBOM"
python scripts/link_workspace_source.py --workspace "Scotia Talent Journey Based Operating Model" --name "JBOM Agile Support"

SharePoint link creation: When copying content from OneDrive, local paths are not shareable. Configure sharepoint_mapping.json (OneDrive prefix → SharePoint base URL). Convert will auto-inject SharePoint URLs so links work for anyone. If you run convert with OneDrive content but no mapping, you'll get a warning with instructions. Add the mapping with add_sharepoint_mapping.py --prefix "OneDrive - X" --base "<paste_url_from_browser>" (paste any file URL; script derives the base).

See content/script-invocation.md for script usage; content/rag-retrieval.md for semantic search flow.

Build

cd skills/ace-context-to-memory
python scripts/build.py