Setting Up AI Search
AI Search sends your episodes and your listeners’ searches to your own account with an AI provider. This page covers connecting that account and every setting on Benecaster → Settings → AI Search. One provider and API key serve all your shows, and each show switches AI Search on separately.
The AI Search settings appear while AI Search is switched on for at least one show (Benecaster → Settings → Add-ons → the AI Search card → Active on).
Choosing a Provider
Under Provider, choose OpenAI, Google Gemini, Cohere or Anthropic. The line under it says which way that provider searches:
- OpenAI, Google Gemini and Cohere build a search index of your episodes ahead of time. Indexing costs a little once per episode; each search after that costs very little. Transcripts are searched, and related episodes by meaning work.
- Anthropic reads your episode list on every search. Nothing is indexed ahead of time, and each search costs more than with an indexing provider. Transcripts are not searched, and related-episode lists set to Related by meaning (AI Search) show your latest episodes.
Model lets you choose between a provider’s models:
| Provider | Models |
|---|---|
| OpenAI | text-embedding-3-small (the default) or text-embedding-3-large |
| Google Gemini | One model, gemini-embedding-001, so there is no Model choice |
| Cohere | embed-english-v3.0 (the default), or embed-multilingual-v3.0 for a show not in English. Cohere reads shorter pieces of text than OpenAI or Gemini, so a long transcript takes more calls to index. A free trial key works but is limited to 40 calls a minute, so indexing a large back catalog on one is slow |
| Anthropic | Claude Haiku 5.5 (the default): The lowest cost per search, and it ranks well for most shows. Recommended. Claude Sonnet 5.5: About 20 times the cost of Haiku per search, for better ranking on a large or varied back catalog. Claude Opus 5.5: About 40 times the cost of Haiku per search, for the best ranking. Repeat searches within five minutes cost much less. |
The lists above are where every site starts, and Benecaster keeps them current. When a provider releases a model Benecaster supports, it can appear in your Model list without an add-on update, with an English name and description; the models above keep their own wording. When a provider retires a model, it leaves the list. If it was your model, AI Search switches to the provider’s default and says so on this screen (The model you chose, {model}, is no longer available, so AI Search is using {model}. Choose a model and save to clear this message.). If the retired model was the default, the first model still offered becomes the default.
Switching provider, or between an indexing provider’s models, re-indexes every episode in the background, with no prompt. Results fill back in as it runs. Saving the same model again re-indexes nothing.
With Anthropic, each search sends your show’s whole episode list, and Anthropic keeps that list cached on its side for five minutes. A search on the same show within five minutes of the last one pays a much lower rate for the list; the first search in five minutes pays slightly more than an uncached search. A search returns at most 20 episodes. On Sonnet and Opus, a search Anthropic’s safety checks decline is run again on another Anthropic model, at that model’s rates; on Haiku it returns no results. If Anthropic retires the model you chose, AI Search switches by itself to the first Anthropic model still offered (Claude Haiku 5.5 unless that is retired too), and this page says The model you chose, {model}, is no longer available, so AI Search is using {model}. Choose a model and save to clear this message.
You pay your provider directly, at their published prices. Each provider’s own site has its current prices.
Connecting Your Provider
- Go to Benecaster → Settings → AI Search.
- Under Provider, choose your provider, and under Model, its model.
- Paste your provider’s key into API key. See Setting Up Your AI Search API Key if you don’t have one yet.
- Click Test connection. On success it reads Connected. {Provider} accepted this key. Save settings to keep it. Testing a key is free with every provider and searches nothing.
- Click Save settings.
How Your Key Is Kept
- A key your provider refuses is never saved. The message reads {Provider} did not accept the API key. Check it, or replace it with a new one.
- A key that could not be checked, because the provider was unreachable or busy, is saved but shown as Saved, but not accepted yet. Use Test connection. until a test succeeds.
- Once saved, the key is never shown again. The field shows a row of dots with Replace and Remove, and the status line reads Connected. Last checked {date}.
- Changing provider asks for that provider’s key: Enter your {Provider} API key. Your current key belongs to another provider and will be replaced.
- Remove asks Remove your API key? AI Search stops working on every show until you add one. Choose Yes, remove it or Keep it.
- If Benecaster can no longer read the key, which can happen after your site’s encryption key changes, the status line reads Benecaster can no longer read your saved API key, so it needs to be entered again. Enter it again and save.
The key is stored encrypted on your site and is never sent to benecaster.com.
The Setup Reminder
Until a provider has accepted a saved key, the dashboard’s Needs Attention list shows Connect your AI Search provider. with Add API key →, which opens these settings. It goes away once Test connection or Save settings confirms the key, and comes back if your provider later stops accepting it.
Search Index
With OpenAI, Google Gemini or Cohere and an accepted key, the Search index box shows Indexed N of M episodes., where M is the published episodes of every show with AI Search on, and a status:
- Indexing in the background. This page updates as it runs.
- Re-indexing in the background for the new provider or model. Results fill back in as it runs. After a switch, the count starts again from 0.
- Up to date. Last pass finished {date}.
- Not indexed yet.
If a batch fails, the box says The last batch could not be indexed: {the provider’s reason} It will be tried again. If your provider refused some episodes’ text, it says N episodes could not be indexed, and are not searchable until they are. The provider said: {reason} Re-index all tries them again.
Guests are indexed too. With Guest Manager on for a show, each guest’s bio is indexed in the background, once for the whole roster and again whenever a guest is saved, so the search panel can find guests by meaning. With Transcription Service as well, each turn a guest takes in a transcript, once the speaker is matched to the guest, is indexed on its own for What a Guest Said. Neither changes the episode count. What this sends to your provider is on Privacy and Data Handling.
Re-index all checks every published episode and sends only text that is new or changed since it was indexed, so unchanged episodes cost nothing. How indexing follows your edits is on AI Search Overview.
Related Episodes
The Related episodes box covers the Related by meaning (AI Search) choice for a show’s Related Episodes list (Which episodes to list, on the show’s Templates tab). Lists are worked out in the background from your search index, so showing them costs nothing. Under Anthropic the box says Related episodes by meaning need OpenAI, Google Gemini or Cohere. With Anthropic, a Related Episodes list set to “Related by meaning (AI Search)” shows your latest episodes instead., and has no settings.
With OpenAI, Google Gemini or Cohere it opens with Related lists ready for N of M episodes. and a status:
- Working them out in the background. Pages without a list yet show your latest episodes.
- Up to date. Last worked out {date}.
- Not worked out yet. They start once the search index is built.
When something stops lists refreshing, it says so: Related lists could not be worked out: {reason} Pages show your latest episodes until they are, and it will be tried again.; or, when your provider stops accepting your key or can’t be reached, that new and edited episodes’ lists are waiting while lists already worked out keep showing. Listeners never see a failure: the page shows your latest episodes.
Keep each list for: 1 day, 1 week (recommended) (the default), 1 month or Until episodes change or I clear them. After this, a list is worked out again the next time its episode page is viewed. Lists are also worked out again whenever an episode of the show is published, edited or unpublished, whatever you choose.
Prefer episodes close in date, Prefer episodes from the same season and Prefer episodes with the same guest: each No (the default), A little or Strongly. Meaning always counts most; these settle close calls between episodes similar in subject. Same season reads each episode’s Season field, so an episode without one is never lifted. Changing any of them works every list out again in the background, at no cost. Restore defaults sets all three back to No and leaves Keep each list for alone.
Prefer episodes with the same guest needs the Guest Manager add-on. Without it the box says “Prefer episodes with the same guest” needs the Guest Manager add-on. With it, two episodes share a guest when Guest Manager links the same guest to both, and an episode sharing at least one guest with the one being viewed is lifted (once, however many it shares). Lists are worked out again whenever a guest or an episode’s guests change, or Guest Manager is switched on or off for a show, still at no cost.
Each choice saves as soon as you change it, and says Saved.
Clear related lists works every list out again in the background, from your search index, at no cost. Until an episode’s list is ready, its page lists your latest episodes.
Who Can Use the Search Box
AI Search is for: Everyone (the default) or Signed-in visitors only. It saves as soon as you choose, and covers every search box on the site.
With Signed-in visitors only, anyone signed in to your site gets the AI search box. A visitor who is not signed in gets Benecaster’s free keyword search box in its place, for the same show, with results as they type. Their searches never reach your provider, and count toward neither the 10-a-minute limit nor the daily limit. A box that was already on screen, or a page your cache saved from a signed-in visit, answers a signed-out visitor Sign in to search.
Use it to keep bots off your provider bill. It changes nothing in your own admin search panel.
Daily Limit on Paid Searches
Each search your provider answers is billed to your account. This limits how many each show can send in a day.
- Limit: A daily limit per show (recommended) or No limit.
- Paid searches per show per day: a whole number from 1 to 1,000,000. The default is 500.
- Click Save limit.
The limit applies to each show separately, and the day runs midnight to midnight UTC. Only a search your provider is asked to answer counts. The same words searched again soon after are answered from a saved copy at no cost: for a day with OpenAI, Google Gemini or Cohere, and for an hour with Anthropic. Your own searches in the admin search panel never count.
Once a show reaches the limit, its search box shows no results to new searches until midnight UTC; pinned results and searches already made that day still show. The box lists each show that reached the limit today: {show} reached today’s limit at {time}. Its search box shows no results until midnight UTC. Raising the limit, or choosing No limit, lets that show search again at once.
Choosing No limit shows With no limit, each visitor is still held to 10 searches a minute, but a bot spread over many addresses could run up your provider bill.
Add Search to Your Site
The Add search to your site box, at the foot of the page, holds the shortcode ready to copy: click it to select it. Put it on any page or post to give your listeners an AI Search box:
[benecaster_ai_search]
Each box can choose its own options, inside the brackets:
show_id="12"— which show to search. Without it, the box searches the show or episode page it is on, or your only show.placement="sidebar"orplacement="modal"— a compact box for a sidebar, or a button that opens search in a pop-up. Without it, a full-width box.result_visibility="locked"— show episodes a visitor can’t open as locked cards with a link to subscribe. Without it, those episodes are left out. See Locked Episodes in AI Search Results.
What the box does is on AI Search Overview.
The Setup Wizard Step
When Benecaster’s setup wizard runs while AI Search is switched on for at least one show, a step headed Connect your search API. appears just before the last screen: AI Search uses your own API account for semantic search. Costs are typically very low at podcast scale. It has Provider and API key, links to Get a {provider} API key, {provider} prices and Step-by-step: getting a key from each provider, and Test connection.
Test connection checks the key at no charge and, once it passes, saves the provider and key, the same values Settings → AI Search shows. Continue → stays unavailable until then, or is available at once when a key is already accepted. The provider gets its default model; choose another on Settings → AI Search. To skip, choose I’ll configure this in Settings.
If you finished setup before switching AI Search on, Benecaster invites you instead: An add-on has a new setup step for {show}., listing Connect AI Search, with Set It Up and No thanks. Set It Up runs just this step. It asks once, on your lowest-numbered show with AI Search on, because the key serves every show, and only while no accepted key is saved. Every answer is final. Saving a key your provider accepts on Settings → AI Search removes the invitation. See Setup Steps From Add-ons You Install Later.
On the Dashboard’s Finish Setting Up list, Connect AI Search shows on every show with AI Search on until your provider accepts a key, even after No thanks. Once a key is accepted, it leaves every show’s list, including a show you switch AI Search on for later. The invitation itself still appears on one show only.
These settings are always available on Settings → AI Search, whatever you answered.
Removing AI Search
Deactivating AI Search on the Plugins screen removes nothing. Deleting it there removes everything it stored: your API key, the provider and model settings, the search index, the related-episode lists, every show’s pinned results and any scheduled indexing. Benecaster’s Reset Plugin Data also clears the search index, the related-episode lists and the pinned results; indexing then starts again by itself.
Next Steps
- Put the search box on your site: AI Search Overview.
- Decide how each box treats episodes a visitor can’t open: Locked Episodes in AI Search Results.
- Search your own back catalog: Searching Your Own Episodes.
See Also
Need this built rather than just documented? See our services →
Last updated