RAG Search (Retrieval-Augmented Generation) lets Maya answer questions using your own documents — call scripts, SOPs, policy PDFs, training manuals, escalation guides, and FAQs. Instead of guessing, Maya retrieves the exact section from your document and presents it. This guide covers how to prepare, format, and upload your content.
What RAG Does for Your Team Included free in all plans
Agent asks "What's the refund policy?" — Maya retrieves the exact clause
New hire asks "What's the opening spiel for account X?" — Maya returns it instantly
Supervisor asks "What's the escalation path for billing disputes?" — Maya finds the SOP step
HR asks "What are the leave entitlement rules?" — Maya pulls the exact policy
Works across WhatsApp, Viber, LINE, and website chat simultaneously
Responds in the language the question was asked in (30 languages)
Speed: RAG search returns results in under 50ms. The bottleneck is GPT response generation — typically 1–3 seconds total for a full answer.
What Documents Work Best Content that answers real questions
Call Center Operations
Call scripts · Opening spiels · Closing spiels · Hold scripts · Transfer scripts · Call flow diagrams (text only) · Escalation trees · Objection handling guides
HR and Policy
Employee handbooks · Attendance and leave policies · Code of conduct · Disciplinary procedures · Probation terms · Benefits summary · Onboarding checklists
Training and Knowledge
Product knowledge manuals · Service descriptions · FAQ documents · Troubleshooting guides · Process SOPs · Compliance requirements · Quality assurance rubrics
Client and Account-Specific
Account-specific guidelines · Client-provided scripts · Account SOPs · Pricing sheets · Product catalogs · Client contact lists (internal use only)
Do not upload: Personal employee records, individual performance data, private financial records, or documents containing passwords or system credentials. RAG content is accessible to anyone Maya speaks with on your channels.
Supported File Formats Text extraction quality varies
Best practice: One document per topic area. A 20-page combined handbook is harder to search than 5 focused 4-page documents (Leave Policy / Attendance / Code of Conduct / Benefits / Onboarding).
Preparing Your Documents for Upload Before you send to your deployment partner
Step 1
Identify what questions your team asks most often
Ask your team leads: "What are the top 10 questions new agents get wrong?" or "What do agents look up most in the first month?" These questions drive what documents matter most. Start with those — not a full document dump.
Step 2
Structure each document with clear section headings
Use plain English section headings that match how people ask questions. Instead of "Section 4.2.1 — Tier B Escalation Protocol," use "When to escalate a billing complaint." Maya searches by meaning, not by document structure — but clear headings improve chunk boundaries and retrieval precision.
Step 3
Remove non-text content
Diagrams, logos, and signature blocks add no search value. Remove them from the version you upload. If a process is documented as a flowchart, rewrite it as numbered steps in plain text. Maya cannot read images inside documents.
Step 4
Name files descriptively
The filename is used as the document source label in Maya's responses. Use names like call-center-escalation-guide.md or hr-leave-policy-2026.txt. Avoid names like document-final-v3-REVISED.docx.
Uploading Your Documents Via Maintenance Audit access
Step 1
Use the direct upload link provided by your deployment partner
The Maintenance Audit add-on includes a direct link to upload content to Maya's document index. Your deployment partner provides this link during onboarding. If you are on a bundle or paid plan, this is included — contact your partner if you have not received it.
Step 2
Upload one document at a time
Each document is split into chunks, embedded, and indexed into the vector database. Large documents take 30–60 seconds to process. Upload, wait for the confirmation message, then upload the next document. Do not close the page mid-upload.
Step 3
Verify the upload completed
After uploading, test by asking Maya a question that requires information from that document. If Maya retrieves the correct information, the document is indexed correctly. If Maya says she does not have that information, contact your deployment partner — the ingestion may need to be re-run.
Storage included: Your plan includes document storage in Supabase. Standard corpora (up to 1 GB) are covered. If your document set is larger — for example, a full product catalog or a legal library — contact your deployment partner to discuss corpus sizing.
Testing Your RAG Search Verify before going live
Step 1
Ask a specific factual question from your document
Pick a specific fact from a document you uploaded — for example, the exact number of sick leave days per year, or the exact wording of your opening spiel for a specific account. Ask Maya in a natural sentence: "How many sick days do full-time agents get per year?" If Maya returns the correct answer with the source document name, RAG is working.
Step 2
Test with a question Maya should NOT be able to answer
Ask about something you have not uploaded — for example, a policy from a different department, or a topic not in any of your documents. Maya should say she does not have that information and ask what else she can help with. If she invents an answer, contact your deployment partner — the system prompt may need a boundary instruction added.
Step 3
Test in each language your team uses
Ask the same question in Filipino, Thai, Hindi, or whichever languages your agents use. Maya retrieves from the same document index but responds in the language of the question. Verify that the retrieved content is correct and the language of the answer is accurate.
Step 4
Ask a question that spans two documents
For example: "What is the escalation path for a billing complaint and what is the refund policy?" Maya retrieves from both the escalation guide and the refund policy document and combines them in a single response. Verify both sources are cited correctly.
Updating and Replacing Documents When policies change
Delete the old document first, then upload the new version — do not upload duplicates
Re-name the file with the new year or version (hr-leave-policy-2027.txt)
Test the updated content before announcing to your team
Inform your deployment partner when you replace a core document — they may need to update the system prompt if a key policy name changed
Stale documents: If a policy has changed but the old document is still in the index, Maya will answer with the outdated information. Regularly audit your document set — especially leave policies, escalation procedures, and account-specific scripts — when they are revised.
This guide covers setting up the RAG Search Engine — Retrieval-Augmented Generation — so Maya answers questions instantly from your own documents rather than general knowledge. The four document types that work best are call center operations content (scripts, spiels, escalation trees, objection guides), HR and policy documents (employee handbooks, leave policies, code of conduct, onboarding materials), training and knowledge base content (product manuals, FAQs, SOPs, QA rubrics), and client or account-specific materials (account guidelines, tailored scripts, pricing sheets). File format ratings: plain text .txt and Markdown .md files give the most accurate retrieval because they contain no formatting overhead; PDF and .docx are good if they are digitally created rather than scanned. Avoid .xlsx and .pptx — table cells and slides are not suitable for semantic chunking. Before uploading, add clear English section headings and use descriptive filenames, because the filename appears in Maya's source citation on every answer. Upload documents one at a time via the direct link provided in the Maintenance Audit add-on — chunking and vector embedding take 30 to 60 seconds per document. After uploading, test with a specific factual question (to confirm retrieval), an out-of-scope question (to confirm Maya admits limited knowledge), a question in a second language (to confirm 30-language response), and a question that spans two documents (to confirm cross-document retrieval). When a document is revised, delete the old version first before uploading the new one to prevent conflicting answers.