# Ensuring Accuracy and Preventing Hallucinations
Odin doesn’t generate guesses - it verifies truths.
Every Odin response is the product of a multi-agent reasoning chain that interprets your question, plans the optimal path to retrieve data, executes deterministic analysis, and validates the results before presenting them back to you.
This architecture makes Odin transparent, auditable, and resistant to hallucinations.
***
## 1. From Question to Insight
When you ask Odin something like:\
\&#xNAN;**“What’s the conversion rate from MQL to SQL for accounts that engaged with Paid Social in the last 60 days?”**
Odin doesn’t respond with a language-model estimate - it executes a multi-step reasoning process to produce a verifiable answer.
This visual illustrates how Odin decomposes a natural language query into structured logic - identifying entities (MQLs, SQLs, Paid Social), mapping them into company-level cohorts, and generating a fully auditable result with defined metrics and methodology.
The outcome isn’t a guess - it’s a computed answer that’s traceable to the underlying data.
***
## 2. Odin: Multi-Agent Orchestration for Data Analysis
Odin’s reliability comes from its **multi-agent architecture**, where each specialized agent handles a precise stage in the reasoning chain.
### **A. Understanding the Question**
When you ask Odin a question - for example:
> *“What’s the conversion rate from MQL to SQL for accounts that engaged with Paid Social in the last 60 days?”*
Odin doesn’t respond with text-based speculation. It interprets your intent using contextual information about your company’s GTM model, definitions, filters, and roles.
This interpretation phase maps:
* **Key entities** (MQLs, SQLs, Paid Social, time windows)
* **Business definitions** (how your org defines funnel stages or engagement)
* **Analytical structure** (what needs to be calculated and grouped)
The result is a **multi-step execution plan** that Odin will follow to generate a verified, reproducible answer.
### B. Generating the Plan
Once Odin parses your question, it constructs a structured plan made of focused analytical steps. Each step produces a verified output that feeds the next. Reasoning and generation are separated. Every step is executed, validated, and logged before the workflow proceeds. That separation is what prevents hallucination.
#### 1) Define entities and scope
Odin identifies the components of the request, the filters, and the grouping logic.\
Example: conversion stages, the engagement source, the time window, and whether results should roll up at the account level.\
This becomes the blueprint for the workflow.
#### 2) Gather governed inputs
Odin retrieves the required stage data and cohorts that match your definitions. Inputs are aligned to the same filters and time ranges to avoid context drift.
#### 3) Segment and contextualize
Odin isolates the relevant audience and time period, then prepares the precise subset needed for the calculation. This ensures the calculation is run on exactly the records the question implies.
#### 4) Compute deterministically
Metrics are calculated with executable code on structured data, not with probabilistic text prediction.\
Example: Conversion Rate = SQL count divided by MQL count within the same scope.
#### 5) Validate and supervise
Results are checked against source data and the original plan. If anything diverges, the workflow is automatically adjusted and re-run until the outputs reconcile.
#### 6) Assemble the answer
Only validated results are presented. The final response includes the metric, the scope that produced it, and links to inspect the underlying data.
### The Outcome: Structural Immunity to Hallucination
Odin’s multi-agent architecture doesn’t just make insights faster — it makes them **trustworthy by design.**
By separating planning, execution, and evaluation, Odin enforces correctness at every stage.
Hallucinations can’t occur because:
1. Every computation must pass verification.
2. Every agent’s output is grounded in source-of-truth data.
3. Every reasoning path is supervised and auditable.
Odin doesn’t speculate - it proves.
***
## 3. Accuracy by Design
Odin’s safeguards make hallucination structurally impossible:
* **Deterministic Computation:** Results are produced by executed code, not generative text.
* **Evaluation Pipelines:** Every answer is cross-checked against its source data before display.
* **Multi-Agent Supervision:** The Supervisor enforces logical and numerical consistency.
* **User Audit-ability:** Each response can be traced back to its dataset, logic, and evaluation chain.
The result is an AI that’s not just accurate - it’s explainable.
***
## Final Thought: AI That Explains Its Reasoning
Odin doesn’t just provide answers - it shows its reasoning and thought process.
Through agent-based orchestration, executable computation, and validation at every layer, Odin turns natural language questions into transparent, verified insights.
It’s AI that earns your trust by proving its accuracy, step by step.
---
# Agent Instructions: Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.
Perform an HTTP GET request on the current page URL with the `ask` query parameter:
```
GET https://docs.hockeystack.com/marketing-intelligence/odin-ai/ensuring-accuracy-and-preventing-hallucinations-1.md?ask=
```
The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.
Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.