I'm building an internal copilot over SharePoint and Google Drive. What retrieval setup should I use?

For an internal copilot over SharePoint + Google Drive, I’d recommend a hybrid, source-aware retrieval architecture rather than “one vector index for everything.”

Recommended setup

1) Ingest into a unified document pipeline

Pull content from both sources into a common schema:

• doc_id
• source = SharePoint / Google Drive
• source_url
• title
• owner
• created_at, updated_at
• permissions / ACLs
• folder/path
• content
• content_type (doc, ppt, pdf, sheet, etc.)
• version

This makes downstream retrieval much easier.

---

2) Use hybrid retrieval

Use both:

• BM25 / keyword search for exact terms, project names, acronyms, policy IDs
• Vector search for semantic similarity

For internal enterprise content, hybrid search is usually much better than pure embeddings alone.

---

3) Chunk documents intelligently

Split by structure, not just fixed length:

• Headings
• Paragraphs
• Bullet lists
• Table sections

Good defaults:

• chunk size: ~300–800 tokens
• overlap: ~50–150 tokens
• preserve document hierarchy and section titles in each chunk

For PowerPoints and docs, include slide/page context in the chunk text.
For spreadsheets, index row groups or sheet summaries rather than raw cells only.

---

4) Store metadata separately and filter hard

Before semantic retrieval, enforce permissions filtering:

• user can only retrieve documents they have access to
• filter by source, team, department, date, doc type if needed

This is critical for SharePoint/Drive because ACLs vary by file/folder.

---

5) Use a two-stage retrieval flow

A strong pattern is:

1. Candidate retrieval

   • hybrid search over chunks
   • permission-filtered
   • fetch top 50–200 candidates

2. Reranking

   • use a cross-encoder / reranker
   • rerank to top 5–15 chunks

This improves relevance a lot, especially for enterprise docs with similar wording.

---

6) Consider document-level + chunk-level indexes

Use both:

• Document-level index for finding the right file
• Chunk-level index for finding the right passage

This helps when:

• user asks “What is the latest vacation policy?”
• you need to identify the policy doc first, then extract the relevant section

---

7) Add freshness-aware ranking

For internal content, recency often matters.

Boost:

• newer versions
• recently modified docs
• canonical docs (policies, handbooks, source-of-truth pages)

But don’t let recency override authority if a policy doc is older but canonical.

---

8) Include source attribution in answers

Always return:

• doc title
• source platform
• link
• quoted passage or referenced section

This builds trust and makes it easy for users to verify.

---

Best-practice architecture

If I were designing this today, I’d use:

• Unified ingestion + normalization
• ACL-aware metadata filtering
• Hybrid retrieval (BM25 + embeddings)
• Cross-encoder reranking
• Doc + chunk indexes
• Freshness / authority boosting
• Citations in final output

---

Practical stack options

If you want a managed approach

• Elastic / OpenSearch hybrid search
• Azure AI Search if you’re already in Microsoft ecosystem
• Vertex AI Search if you’re more on Google Cloud

If you want to build it yourself

• Ingestion: custom connectors to SharePoint + Google Drive
• Index: Elasticsearch/OpenSearch + vector index, or a vector DB plus keyword engine
• Reranker: sentence-transformers cross-encoder or hosted rerank API
• Orchestration: LlamaIndex / LangChain / custom retrieval service

---

Important SharePoint/Drive specifics

SharePoint

• Watch for nested permissions inherited from sites, libraries, folders, files
• Preserve version history where useful
• Handle Office docs with embedded tables and comments carefully

Google Drive

• Use file ACLs and shared drive memberships
• Be careful with shortcuts, duplicates, and “latest” file versions
• Extract text from Docs/Slides/Sheets properly, not just exported PDFs

---

My default recommendation

If you want the most reliable setup:

Permission-filtered hybrid search + chunk-level indexing + reranking + citations

That’s the sweet spot for enterprise internal copilots.

If you want, I can also propose:

1. a reference architecture diagram,
2. a chunking/indexing strategy by file type, or
3. a minimal production-ready stack for your cloud environment.