DocuChat AI

Admin — Knowledge Base

• Upload: Click "Upload" → drag-and-drop or select files. You can drop multiple files at once, or drop additional files to add to the selection. Supports PDF, DOCX, TXT, Markdown, Excel (.xlsx/.xls), CSV, SQLite (.sqlite/.db/.sqlite3), and SQL dumps (.sql).
• Database files: Upload your blog or CMS SQLite database directly — all tables are extracted and indexed as searchable text.
• Add URL: Paste a public website URL to scrape and index its content.
• Batch delete: Use the checkboxes to select multiple documents, then click "Delete Selected" to remove them all at once.
• Enable/Disable Sources: Toggle any document on/off. Disabled sources are excluded from chat answers instantly — no re-processing needed.
• Re-embed All: After changing the embedding model, click to regenerate all embeddings without re-uploading.
• Retry: If processing failed (network issue), click Retry on the failed document.
• Re-upload: Same filename replaces the old version automatically.
• Status: Documents show "processing" with a progress bar that auto-refreshes until "ready". Do not ask questions until all documents show "ready".
• Format guidance: PDF, Word, and Markdown work best for SOPs and narrative content. Excel/CSV for reference data. SQLite/SQL for database exports (blogs, CMS content).

Admin — Chat Logs

• View all chats from all users — even after users clear their own conversations.
• Click a conversation to expand and see all messages.
• Export as CSV for analysis (FAQ identification, training needs).
• Clear All Logs permanently removes all history (with confirmation).

Admin — AI Configuration Guide

This section explains every setting in the AI Configuration panel. Only Tech Admins can change these. Each question a user asks costs a small amount — understanding these settings helps control quality and cost.

What is an API Key?

An API key is like a password that lets this app talk to an AI service (OpenAI, Google, etc). You get it from the AI provider's website. It's tied to your billing account — every question users ask here costs a tiny amount on your account. Typical cost: $0.01–$0.05 per question with cheap models, $0.05–$0.30 with premium models.

LLM Provider (Chat AI)

This is the company whose AI generates the chat answers. Options:

• OpenAI — Most popular. Models: gpt-4o-mini (cheap, fast, recommended), gpt-4o (smarter, 10x more expensive). Get key from platform.openai.com.
• Google Gemini — Has a free tier. Models: gemini-1.5-flash (fast, cheap), gemini-1.5-pro (smarter). Get key from aistudio.google.com.
• Anthropic Claude — Known for careful, accurate answers. Models: claude-3-haiku (cheap), claude-3.5-sonnet (premium). Get key from console.anthropic.com.
• OpenRouter — A middleman that gives access to all models through one key. Useful if you want to try different models without creating accounts everywhere. Get key from openrouter.ai.
• Local (Ollama) — Runs AI on your own computer. Free (no API cost) but requires a powerful machine with a good GPU. Quality varies by model.

Model

Each provider offers multiple models. Think of models like car engines — some are economy (cheap, fast) and some are premium (smarter, slower, expensive). For document Q&A, economy models work great because the answer comes from your documents, not the AI's general knowledge. Start with the cheapest model and only upgrade if answers feel inadequate.

Embedding Model

This is separate from the chat model. The embedding model converts your documents into a searchable format (numbers that represent meaning). It runs once when you upload a document, and once per question to match your question to relevant sections.

• text-embedding-3-small — Recommended. Very cheap ($0.02 per million tokens ≈ pennies for thousands of documents). Good accuracy.
• text-embedding-3-large — 6x more expensive but slightly more accurate for nuanced content. Usually not needed.

Warning: If you change the embedding model after uploading documents, you must click "Re-embed All" to regenerate all document embeddings. This costs API credits proportional to your total document content.

Temperature

Controls how "creative" vs "precise" the AI is. Scale: 0.0 (robotic, exact) to 1.0 (creative, unpredictable).

• 0.0–0.2 — Best for document Q&A. Gives factual, consistent answers from your content.
• 0.3–0.5 — Slightly more natural language, still mostly factual.
• 0.6–1.0 — More creative, may paraphrase heavily or add opinions. Not recommended for factual document queries.

No cost impact — same price regardless of temperature.

Max Chunks

When a user asks a question, the system finds the most relevant sections ("chunks") from your documents and sends them to the AI as context. This setting controls how many sections are sent.

• 3–5 — Good for simple questions. Cheaper (fewer tokens sent to AI).
• 6–10 — Better for complex questions that span multiple documents or topics. Costs more per question.
• 10+ — Rarely needed. Sends a lot of context, significantly increases cost per question.

Cost impact: Each chunk adds ~200–400 tokens to the AI request. At 5 chunks, a typical question uses ~3,000 tokens. At 10 chunks, it doubles.

Similarity Threshold

Controls how closely a document section must match the question before being included as context. Scale: 0.1 (very loose) to 0.99 (exact match only).

• 0.5–0.6 — Casts a wide net. Finds more results but some may be irrelevant. Good if users phrase questions loosely.
• 0.7 — Recommended default. Good balance of relevance and recall.
• 0.8–0.9 — Very strict. Only returns near-exact matches. May miss relevant content if phrased differently than the document.

No direct cost impact, but lower thresholds may result in more chunks being sent (up to Max Chunks).

Streaming Responses

When enabled, the AI's answer appears word-by-word as it's being generated, like watching someone type. When disabled, the user sees a loading indicator and then the full answer appears at once. Streaming feels faster and more responsive. No cost difference — enable it unless you have a specific reason not to.

System Prompt

The system prompt is a hidden instruction sent to the AI with every question. It tells the AI how to behave — like giving an employee their job description. The default prompt says: "Only answer from the provided document context. If the information isn't in the documents, say so. Include citations."

You can customize it to change the AI's personality, add formatting rules, or adjust how it handles conflicts between documents. The prompt itself adds a small fixed cost to every question (typically 100–300 tokens).

Cost Summary

Total cost per question = system prompt tokens + document chunks tokens + user question + AI answer. With gpt-4o-mini and 5 chunks, expect roughly $0.01–$0.03 per question. A team of 10 people asking 20 questions/day ≈ $2–$6/month. Premium models (gpt-4o, claude-3.5-sonnet) cost 5–15x more.

Admin — Other Settings

• Suggested Prompts: Up to 10 prompts shown as cards on empty chat screens. Users click one to pre-fill the message box. Drag to reorder.
• User Management: Create users (username + display name + password + role). Change roles. Remove users.
• Chat Usage Chart: Bar chart showing messages per user, with date range selector.

Admin — User Roles

• User: Chat only.
• Admin: Chat + Documents + Chat Logs + Admin (except AI config/prompt).
• Tech Admin: Everything including AI configuration and system prompt.

Tech Admins can promote/demote anyone. Admins can only toggle between User and Admin. Users cannot change roles.

Admin — Troubleshooting

• "I couldn't find that information" — Document may not be uploaded, or question phrasing is too different. Rephrase or check document status.
• Document "error" + "fetch failed" — Network/VPN blocking external API. Disconnect VPN, then Retry.
• Wrong page numbers — Re-embed the document to regenerate page-aware chunks.
• Slow responses — Enable streaming. Use a faster model (gpt-4o-mini).
• URL add failing on deployed app — Check server can reach external sites. Browser cache may serve old JS — bump version or hard refresh.
• Session lost after restart — Expected. Sessions are in memory. Users re-login.

Admin — Data & Backup

• Database: data/chatbot.db — all data (users, embeddings, config, logs).
• Files: uploads/ — original uploaded documents.
• Backup: Copy both to preserve everything.
• Deploy: Copy project folder, run npm install, start with pm2 or node src/index.js.