--- title: Get vault recommendations description: Find idle funds and rank vaults, by scanning your wallet or by describing a goal. url: https://pr-1-9bee85ef9c41.thally.app/guides/ai-concierge --- # Get vault recommendations Find idle funds and rank vaults, by scanning your wallet or by describing a goal. The Yield Concierge answers two different questions. **Idle sweep** starts from what you hold: "what should I do with this?" **Intent** starts from what you want: "find me a safe USDC vault above 5% on Arbitrum". Both live under **Put idle assets to work** on the **My Positions** tab of LI.FI Earn. Neither needs a connected wallet to *look* — paste any address and it will scan that one instead. ## Idle sweep Hexkit scans your balances across every chain that has vaults, finds tokens sitting idle, and suggests where each one could go. #### Let it scan A status line reads `12 of 14 chains · 8 idle · 0 selected` as it works. Tokens worth under $1.50 are collapsed into a "low-value tokens hidden" group — click to expand. #### Select what to deploy Tick the assets you want to use. Each has a **0–100% slider**, so you can deploy part of a balance and keep the rest. #### Choose how to route **Per-asset** sends each token to its own vault, possibly on different chains. **Consolidate** brings everything into one vault, bridging as needed. #### Review the picks Each asset gets up to four suggestions — **Best**, **Safest**, and two alternatives. Click a pill to change the destination. The queue builds and starts automatically once you have selections with a destination. There's no separate Run button — but nothing is signed until you approve it in your wallet. > **Note:** If the scan fails, there's no rescan button. Reload the page. ## Intent Describe what you want in plain English: > Put my USDC into the safest vault above 5% APY on Arbitrum Press **Get recommendations** and Hexkit turns that into structured filters, shown as removable **Parsed filters** chips — the objective, token, chain, APY floor, TVL floor, and so on. Remove a chip to loosen that constraint; the shortlist re-ranks without re-reading your sentence. Below the list you'll see how the universe was narrowed, like `Picked from 24 intent-filtered vault(s) (of 812 total).` When nothing matches, it tells you exactly what was dropped — wrong token, below APY floor, excluded protocol. If your token has no matching vault, Hexkit relaxes that one filter and says so: *"No PEPE vaults matched your criteria. Showing top vaults reachable via swap."* ## How vaults are ranked Filtering happens locally first. Vaults are dropped for non-transactional status, chain or token mismatch, APY outside your bounds, TVL below your floor, and protocol exclusions. Two risk filters also apply: - **High risk** — APY above 250% *and* TVL under $2M - **Caution** — TVL under $250K, reward-heavy APY, a 1-day spike over 3× the 30-day average, or a declining trend Both are hidden by default with a toggle to reveal them. The survivors go to the model, which picks **Best** and **Safest** (never the same vault) and writes a one-line rationale for each. It's told to prefer sustainable base APY over emissions, to weigh TVL and protocol maturity, and to reject anything whose entry cost takes more than six months to earn back. > **Note:** The model only ever chooses from vaults Hexkit already shortlisted. It cannot invent one — a hallucinated name resolves to nothing rather than into your queue. ## AI or rules Every recommendation card carries a badge: **AI** or **Rules**. **Rules** means the ranking came from a local heuristic — highest APY above a $10M TVL floor for Best, highest TVL for Safest. Its rationales are formulaic and numeric (`Highest APY (7.42%) among 14 eligible vaults.`), which is the giveaway. Fallback happens when the model is disabled, when the call fails, when the response doesn't parse, or when only one candidate survived filtering. > **Warning:** Rules mode has a real limitation in per-asset intent mode: the ranking is computed once and reused, so **every asset can get the same vault**. The differentiation between tokens happens inside the model prompt. If you see one vault suggested for everything, check whether the badge says `Rules`. ## Turning the AI off Set `VITE_LLM_MODE=off` in your own deployment and every recommendation becomes rules-based. Scanning, filtering, the flow diagram, and execution all still work. > **Warning:** In `off` mode, **Intent silently ignores what you typed.** There's no error and no notice — the parsed filters simply show "(no specific criteria — ranking entire vault universe)" and you get a generic list. If your filters vanish, check this setting first. Self-hosters: the model runs on the operator's key, so there's no per-user quota. Your only costs are on-chain — gas, swap fees, bridge fees. > **Note:** Recommendations for a **pasted** address are limited to 3 model calls per hour, and one call is used per selected asset. Hitting it shows the generic *"AI recommender unavailable"* alert with a Retry button that won't help until the hour is up. Your own connected wallet isn't rate-limited. ## Executing a plan The queue runs **one leg at a time**, each rendered as a full deposit flow. You approve every transaction in your wallet — a same-chain leg needs an approve and a deposit; a cross-chain leg adds a bridge step. **Next Step** advances, and stays disabled until the current leg finishes. When everything is done it becomes **Reset**. ### When a leg fails A failed leg offers **Skip & continue**. What that means depends on whether funds are recoverable. If they aren't at risk, skipping just moves on. If they might be — escrowed, or delivered but not yet deposited — you'll get a hint like *"Funds were delivered on the destination chain. Retry the deposit step to finish."* and an amber **Skip & continue (funds may be stuck)** button. > **Warning:** Skipping a recoverable leg **does not recover the funds**, and it removes the retry and refund controls from that step. Your remaining route is the vault drawer, manually. Read the hint before skipping — retrying is usually the right move. Legs that can't run at all are listed as **Skipped** with a reason rather than dropped silently — a non-EVM source chain, a native token where an ERC-20 is required, no route for the pair, or an amount below the dust threshold. > **Note:** Changing a destination while a queue is part-finished rebuilds and restarts it from step one. Set destinations before starting. ## Related - [Deposit into a vault](/guides/deposits) — the deposit flow each leg runs - [How cross-chain works](/concepts/intent-orders) — escrow, deadlines, refunds - [Self-hosting](/self-hosting) — `VITE_LLM_MODE` and the model proxy