Skip to main content
Running into issues? Here’s how to troubleshoot common problems and get FlipsideAI back on track.

​
Context window drift

​
The problem

Long conversations lose memory of earlier messages. FlipsideAI may:
  • Forget details you mentioned earlier
  • Contradict itself
  • Lose track of your original goal

​
Why it happens

AI models have a limited “context window”—a maximum amount of text they can remember. Long chats exceed this limit.

​
Solutions

Periodically ask:
"Summarize what we've found so far in 5 bullet points"
Copy the summary, then start a fresh chat with that context.
Don’t try to tackle multiple unrelated analyses in one conversation. Split them: - Chat 1: Aave TVL analysis - Chat 2: Uniswap liquidity research - Chat 3: Cross-chain bridge volumes
If FlipsideAI seems confused, restate your goal: text "To recap: we're analyzing Arbitrum DEX volumes over Q1 2024, focusing on the top 5 protocols. So far we've looked at Uniswap and Curve. Now let's examine SushiSwap."
Teams let you organize multi-conversation projects without losing context across chats.

​
Hallucinations

​
The problem

FlipsideAI gives you numbers or insights that seem plausible but are incorrect.

​
Why it happens

LLMs can “hallucinate”—generate confident-sounding answers that aren’t grounded in data. This happens when:
  • The model misinterprets a query
  • Data is ambiguous or incomplete
  • The model fills gaps with plausible-sounding guesses

​
How to detect

Watch for red flags: - Suspiciously perfect numbers: Round figures that seem too convenient - Unrealistic scale: TVL of $10B for a tiny protocol - Contradictions: Numbers that don’t match earlier responses - Vague sourcing: No clear query or data table referenced

​
Solutions

Always check:
  • The SQL query used
  • The raw data returned
  • Whether the filters make sense
If something looks off, the query might be wrong.
"Explain how you calculated this TVL number step by step"
Walking through the logic helps catch errors.
Verify against:
  • DeFiLlama (for TVL)
  • CoinGecko/CMC (for prices)
  • Dune Analytics (for volumes)
  • Protocol dashboards
If Flipside’s numbers diverge significantly, investigate why.
"These results seem high. Can you validate using an alternative approach?"
"Cross-check this TVL against DeFiLlama data"
FlipsideAI can recalculate using different methods.

​
Query errors

​
The problem

Queries fail to execute or time out.

​
Common causes

  • Too much data: Query spans too many rows or too long a time period
  • Complex joins: Multiple table joins slow down execution
  • Incorrect table names: Typo or outdated table reference
  • Schema mismatches: Trying to join incompatible columns

​
Solutions

Instead of:
"Show me all Uniswap swaps ever"
Try:
"Show me Uniswap swaps over the past 30 days"
Smaller date ranges = faster queries.
Instead of:
"Show me all DEX activity"
Try:
"Show me DEX activity on Arbitrum for USDC and ETH"
Filtering reduces query size.
Flipside’s ez_* tables (e.g., ez_dex_swaps) are optimized and faster than raw fact tables.
"Use the ethereum.defi.ez_dex_swaps table instead of fact_event_logs"
For complex analyses, break into sequential queries:
"First, show me total DEX volumes.
Then, break down by protocol.
Finally, add time series data."
Sequential queries are more reliable than one massive query.
"That query failed. Can you try using a different table?"
Alternative tables may have the same data with a better schema.

​
Performance issues

​
The problem

Queries are slow or FlipsideAI takes a long time to respond.

​
Solutions

Add explicit limits:
"Show me the top 10 protocols by volume (limit to 10 rows)"
This prevents returning millions of rows.
Filtering by address is faster than filtering by token symbol:
"Analyze transfers for 0xA0b86991c6218b36c1d19D4a2e9Eb0cE3606eB48 (USDC)"
Symbols can be ambiguous; addresses are precise.
For high-level metrics, use daily or hourly aggregates instead of transaction-level data: text "Use the ez_core_metrics_hourly table for Solana stats"
Don’t ask for “everything”:
❌ "Show me all token transfers on Ethereum"
✅ "Show me USDC transfers on Ethereum over the past 7 days"

​
Incorrect or unexpected results

​
The problem

The numbers don’t match your expectations or known benchmarks.

​
Common causes

  • Token vs USD amounts: Confusing raw token balances with USD values
  • Duplicate counting: Rows are joined incorrectly, inflating counts
  • Wrong filters: Date ranges, chains, or protocols don’t match your intent
  • Data gaps: Some chains or protocols may have incomplete historical data

​
Solutions

"Are these values in USD or in token units?"
"Convert all amounts to USD using hourly prices"
Expand the Flipside Data box and check:
  • Filters (WHERE clauses)
  • Joins (JOIN conditions)
  • Aggregations (SUM, COUNT, AVG)
If something looks wrong, point it out:
"This query is joining on the wrong column. It should join on tx_hash, not block_number."
Compare to known benchmarks:
  • Does this TVL match DeFiLlama?
  • Are these volumes consistent with protocol dashboards?
  • Do user counts align with Dune or Token Terminal?
If not, dig into why.
"This TVL seems low. Recalculate including LP tokens and staked assets."

​
Workflows not triggering

​
The problem

You expect a Workflow to activate, but FlipsideAI uses general mode instead.

​
Why it happens

  • Your prompt doesn’t match the Workflow’s trigger keywords
  • The Workflow doesn’t exist for your use case
  • FlipsideAI misinterprets your intent

​
Solutions

Use the @ symbol to browse and select:
@TVL-Analysis Compare Aave and Compound
Include Workflow-specific terms: - For TVL Workflow: “calculate TVL” or “total value locked” - For cohort Workflow: “cohort analysis” or “user retention” - For liquidity Workflow: “liquidity depth” or “LP behavior”
Ask:
"Is there a Workflow for analyzing bridge volumes?"
If not, use general mode or request a new Workflow.

​
Live Query Mode not working

​
The problem

You want real-time data, but FlipsideAI uses historical tables instead.

​
Why it happens

Live Query Mode must be explicitly requested—it’s not enabled by default.

​
Solution

Use explicit language: 
"Use live query to get the current WETH balance for this address"
"Query the node for real-time pool liquidity"
"Fetch live pricing data for ETH"
Keywords: “live query,” “current,” “real-time,” “query the node.”

​
Artifacts not rendering

​
The problem

Artifact generation fails or the artifact displays incorrectly.

​
Common causes

  • Too much data for the visualization
  • Complex interactions causing rendering issues
  • Browser compatibility problems

​
Solutions

"Reduce this to the top 10 data points"
"Remove interactive filtering"
"Use a simpler chart type (line instead of scatter)"
  • Refresh the page - Try Chrome, Firefox, or Edge - Disable browser extensions that might interfere
"Instead of an interactive artifact, generate a static PNG"

​
Data freshness issues

​
The problem

Data seems outdated or doesn’t include the latest activity.

​
Why it happens

Flipside’s tables are updated periodically (hourly, daily, etc.), not in real-time.

​
Solutions

Ask:
"When was the ethereum.defi.ez_dex_swaps table last updated?"
FlipsideAI can tell you data latency.
For truly current data: text "Use live query to get the latest pool state"
Avoid queries that end “today”—use yesterday or earlier:
❌ "Show me activity through today"
✅ "Show me activity through yesterday"

​
Pro tips for avoiding issues

Begin with a basic query, verify it’s correct, then add layers:
  1. “Show me Uniswap volumes” (verify this works)
  2. “Break it down by pool” (add complexity)
  3. “Add time series” (add more complexity)
  • Specify chains: “on Arbitrum” - Specify time periods: “from 2024-01-01 to 2024-03-31” - Specify tokens: “USDC (0xA0b86991…)” - Specify what you want: “in USD, not token units”
Save troubleshooting patterns as Rules: text Rule: "Always validate TVL results against DeFiLlama before finalizing"
If results look off:
"Walk me through how you calculated this"
"What tables did you use?"
"Why did you choose this approach?"
Understanding the methodology helps catch errors.

​
Still stuck?

If you’ve tried these solutions and still can’t resolve the issue:
  1. Simplify your question: Break it into smaller parts
  2. Start a new chat: Context drift may be causing confusion
  3. Try a different approach: Ask the same question in a different way
  4. Check the docs: Other pages in this guide may have specific solutions
  5. Reach out for support: Contact Flipside support with details about the issue

​
Next steps

Data & results

Deep dive into validating and working with data

Analysis modes

Understand how different modes handle queries

Effective prompting

Craft better prompts to avoid issues

Workflows

Use Workflows for complex, reliable analyses