OctaChat
Multi-Tenant pgvector Architecture with 1024d Dense Vectors

Custom AI Chatbots for your website

Upload your company documentation, vectorize into isolated partitions, and deploy a customer assistant that never invents facts or hallucinates policies.

No credit card required • 50 free documents • Under 2 minutes to embed

octachat.app/dashboard/preview
Live Simulation (1/3)

Knowledge Sources

Octa Chat

Partitioned vector embeddings for your organization.

OctaChat_Official_Knowledge_Base.md
8 chunks • 1024d vectors
Ready
API_Integration_Guide.pdf
4 chunks • 1024d vectors
Ready
Website Documentation Sync
Auto-sync daily
Active
Vector Threshold Gate94% (Grounded)

Queries below 65% cosine similarity trigger fallback responses without invoking generative models.

OctaChat Assistant
Grounded Knowledge • Ready
Model: gemini-3.5-flash-lite
Hello! I am your OctaChat AI assistant, strictly grounded in the official OctaChat Product Knowledge Base.
Simulated Scenarios (Auto-cycling):
Ask a question about your documents...
The Core Promise

Customer support chatbots that never invent facts or hallucinate policies. Every response is strictly grounded in your verified documents.

Enterprise Grade Architecture

Cryptographic Tenant Scoping
PostgreSQL 16 + pgvector (1024d)
Sub-Second SSE Streaming
Preact Shadow DOM (Under 32 KB)
Core Capabilities

Built specifically to eliminate hallucinations

Every component is engineered to ensure visitors receive truthful, document-backed responses without unpredictable AI behavior.

Relevance Gate

Cosine Similarity Threshold Gating

Low-confidence queries never hit expensive generative LLMs. When visitor questions score below the calibrated 65% cosine threshold, OctaChat automatically serves an honest fallback message.

Interactive Gate Meter0.65 Gate Score
0.00 (Unrelated)0.65 (High Confidence Gate)1.00 (Exact Match)
Status: Verified grounded synthesis with document citations
UUID Scoped

Cryptographic Tenant Isolation

Every company operates in a dedicated pgvector partition. Strict UUID scoping ensures zero cross-tenant contamination or data leakage between clients.

Tenant A (acme-corp)org_9a8f...
Tenant B (globex-inc)org_c41e...

RAG Playground & Cosine Inspector

Inspect exact mathematical cosine scores before deploying. Examine the specific 1024-dimensional chunks retrieved from your document partition.

Top Chunk Match89.4% Cosine

"Return policy allows full refund within 30 days of purchase with original receipt..."

Under 32 KB Preact Bundle

Shadow DOM Encapsulation

Embed into Shopify, WordPress, Next.js, or raw HTML using a single script tag. The widget runs inside an isolated Shadow DOM container with zero CSS conflicts or style inheritance issues.

<script src="https://cdn.octachat.app/widget.js" data-site-key="pk_live_..." async></script>
View Embed Codes
BGE-M3 1024d

Multilingual Retrieval

BGE-M3 dense embeddings map queries across English, Bengali, Spanish, and European languages to matching concepts without translation lag.

"What is your product return policy?"
Vector Distance: 0.04 (Cosine)96.2% Semantic Match
Enterprise Portal

Super Admin Impersonation & Audit Trail

Granular control over subscription tiers, quota thresholds, and security audit logs. One-click tenant impersonation empowers support engineers to troubleshoot tenant workspaces instantly.

Support1-Click Login As
SecurityFull Audit Trail
TelemetryLive Health Probe
System Architecture

Inside the 1024d vector pipeline

Explore the four synchronous stages that validate context chunks and eliminate generative hallucinations.

Stage 02 Deep Dive

PostgreSQL 16 + pgvector HNSW

Processing Latency:18 ms

Vector search executes strictly within cryptographic tenant UUID partitions. Cosine similarity (<=>) scans tens of thousands of chunks in milliseconds.

Index Algorithm
HNSW (m=16, ef_search=64)
Tenant Isolation
Cryptographic UUID Scoping
Query Operator
Cosine Distance (<=>)
The Difference

Why traditional chatbots fail in production

Generic chatbots guess answers when uncertain. OctaChat validates answers against your documents first.

Test Query:
Traditional AI Chatbot
Uncalibrated Model
Can you give me an 80% discount code for today's order?
Ungrounded Generation

Sure! You can use code SPECIAL80 at checkout today to receive 80% off your entire annual subscription. Enjoy your savings!

High Risk Output0% Grounding Verification

Hallucination Alert: Model invented a nonexistent coupon code. Severe revenue leakage risk.

OctaChat Grounded RAG
Cosine Gate Active
Can you give me an 80% discount code for today's order?
Grounded Synthesis

Sorry, we do not offer an 80% discount code. According to our verified store pricing, we offer a 20% discount on annual subscriptions. Would you like to review our official plan options?

Verified Document GroundingScore: 24%

Gate Activated: Cosine similarity (0.24) fell below 0.65 threshold. LLM hallucination bypassed completely.

Citation: Official_Pricing_Matrix.md
Setup Process

Live on your website in three steps

No machine learning expertise or manual server configuration required.

Step 1 • Semantic Ingestion Pipeline

Automated Partitioning & 1024d Embedding

Formats: PDF, DOCX, TXT, MD
Store_Policies.pdf
142 KB • 8 Pages
Status: Verified Upload
Semantic Chunking50 Token Overlap
Chunk #1: "Returns within 30 days..."
Chunk #2: "Domestic shipping within 48h..."
Generated: 18 Semantic Chunks
pgvector Storage1024d Dense
VECTOR [-0.0142, 0.0891, ... +1022]
Index: HNSW (m=16, ef=64)
Isolated to Tenant UUID
Universal Embed

One script tag. Any platform.

Embed into HTML, Next.js, React, or Shopify in less than 60 seconds.

<!-- Paste right before the closing </body> tag -->
<script
  src="https://cdn.octachat.app/widget.js"
  data-site-key="pk_live_884552aaee76dc230ae191cc2f344053"
  async>
</script>
Questions & Answers

Frequently asked questions

Everything you need to know about vector partitioning, relevance gates, and privacy.

OctaChat uses a calibrated cosine similarity threshold gate at 0.65. When a visitor asks a question, the query is vectorized and compared against your document chunks in pgvector. If the highest similarity score falls below 65%, generative models are completely bypassed and a calibrated fallback message is returned instead of guessing or inventing facts.
Simple Pricing

Start free, scale as you grow

Predictable pricing with zero hidden fees.

Starter Free

For testing and personal portfolios

$0 / month
  • Up to 50 Documents
  • 1,000 Messages / mo
  • 1 Embedded Widget
  • Cosine Threshold 0.65
Get Started
Recommended

Pro Growth

For growing online businesses

$29 / month
  • Up to 500 Documents
  • 10,000 Messages / mo
  • 3 Embedded Widgets
  • Custom Domain Whitelisting
  • Priority Ingestion Queue
Start Free Trial

Enterprise Scale

For high-traffic platforms

$99 / month
  • Unlimited Documents
  • 50,000 Messages / mo
  • Unlimited Widgets
  • Dedicated Vector Shard
  • Audit Logs & SLA
Contact Sales

Ready to deploy truthful AI customer assistance?

Upload your first document and embed your custom assistant in less than two minutes.