INTERNAL KNOWLEDGE AND RAG
RAG Chatbot for Company Documents
A knowledge-base chatbot that answers staff questions from the company's own documents, and keeps its index current as those documents change.
- PROJECT
- RAG Chatbot for Company Documents
- INDUSTRY
- Internal Knowledge and RAG
- STACK
- n8n · Google Drive · Pinecone · Google Gemini · RAG · Vector Database · AI Agents
OVERVIEW
The project
Employees were asking each other questions that company documents already answered, in files nobody could find. Allzone built a chatbot that watches the company's Drive folder, ingests every new or updated document into a vector store, and answers from that material rather than from a model's general training.
CHALLENGE
What made it difficult
A knowledge assistant is only as current as its index. If ingestion is a manual step it drifts out of date the first week nobody runs it, and staff stop trusting the answers shortly after that.
The other failure mode is worse than being out of date. An assistant that answers confidently where the material does not cover the question teaches people to check nothing, so it needed a defined fallback instead of a guess.
WHAT WE DID
The work delivered.
- 01
Document watching and ingestion
The workflow monitors a Google Drive folder for new and updated files, then downloads, splits and embeds each one without anyone triggering it.
- 02
Vector store and retrieval
Recursive chunking with Gemini embeddings into Pinecone, queried through a retrieval tool the agent calls deliberately rather than searching on every turn.
- 03
Grounded answering
An agent that answers from retrieved company material, with conversational memory across a session and an explicit fallback when the documents do not cover the question.
ARCHITECTURE
How it fits together
- File created and file updated triggers feeding one ingestion path, so new and revised documents are handled identically
- Recursive character splitting ahead of embedding, so a retrieved chunk is big enough to answer from and small enough to be specific
- Pinecone as the vector store, reached through a retrieval tool exposed to the agent
- Conversational memory scoped to the session, which is what keeps follow up questions coherent
TECHNOLOGY
The stack
- n8n
- Google Drive
- Pinecone
- Google Gemini
- RAG
- Vector Database
- AI Agents
OUTCOME
What exists now.
Staff get answers from the company's own policies and documentation instead of asking a colleague who may not know either
The index updates itself when a document changes, so the assistant does not quietly go stale
The same pattern carries to HR policies, SOPs, product documentation and support material without being rebuilt each time
NEXT STEP
Tell us what you're building.
Bring us the problem with its real constraints attached. We will tell you what we would build, and what we would not.