Appearance
Image search
Image search lets your agent match a customer-supplied photo against your product images — useful when a shopper sends a picture ("do you have this?") instead of typing a query. MyChatBot compares the photo to your catalog, blends that visual signal with the normal text search, and pushes the right product to the top of the results.
Cheat sheet
| Concept | What it means for you |
|---|---|
| Photo in, product out | A customer sends an image; the agent finds the closest matching products in your catalog. |
| Works with text too | If the customer sends a picture and a description, both are used together for better precision. |
| Top-of-list = best match | A confident visual match is pushed to the first position so it isn't buried under a generic text query. |
| Needs image indexing | Only catalogs indexed with image search turned on can answer photo queries. |
| Show and confirm | The agent should display the top image and ask the customer to confirm — that's the intended verification step. |
How matching works (in plain terms)
When a photo comes in, MyChatBot trims it to isolate the product (cropping out screenshot borders and app chrome) and then looks for a match in three passes, stopping as soon as it's confident:
- Exact copy. Is this literally one of your catalog images (or a light crop/screenshot of it)? If so, that product wins immediately.
- Near-duplicate. No exact copy, so it looks for images that are visually almost identical — the same product shot from a slightly different angle, lightly edited, or re-compressed.
- Visually similar. It compares overall look and composition to separate close-but-different items and pick the best remaining candidate.
The system treats these three kinds of matches with different levels of trust — an exact copy is trusted far more than a "looks similar" match. That trust level is used internally to decide what gets promoted to the top of the list; it is not shown as a score in the results.
What the agent actually gets back
The product search returns a normal list of products — the same shape as a text search. A strong image match shows up as the first product in the list, not as a separate "confidence" number. To verify, have the agent show that top image and ask the customer "Is this the one?"
Deeper visual-similarity matching
Beyond duplicate detection, MyChatBot can also apply a deeper visual-understanding step that matches images by what they look like rather than just pixel structure. This is a per-catalog capability MyChatBot manages for you:
| Deeper matching on | Duplicate-detection only | |
|---|---|---|
| Recall | Higher — finds semantically similar items, not just near-identical photos | Lower — finds copies and near-duplicates only |
| Speed | Slower per query | Faster |
| Trade-off | Can occasionally over-match loosely similar items | Can occasionally confuse items with a similar layout |
| Best for | Catalogs with many visually similar SKUs (e.g. lots of similar dresses) | Latency-sensitive flows where duplicate-detection is enough |
Enabling or disabling deeper matching
Deeper visual matching is a per-catalog setting MyChatBot manages, and changing it requires re-indexing the catalog. If you want it turned on or off, contact MyChatBot support on Telegram (https://t.me/mychatbot_support) with your account and the product integration in question. MyChatBot also tunes overall image relevance per catalog — if matches feel too loose or too strict, support can adjust it.
Combining a photo with text
When a customer gives both a picture and a description, the agent searches with both at once and MyChatBot merges the two result lists. The visual match is weighted as the primary signal, with the text query refining and disambiguating it — so a strong photo match wins even against a broad query like "evening dress" (which matches every dress), while the text still helps break ties.
If the image step fails or finds nothing, the search quietly falls back to text results — image search never blocks a query.
Look-alikes and the top match
Sometimes a different product with a similar layout looks close enough to get promoted to the top — for example an Instagram screenshot of one dress that resembles another. MyChatBot guards against this so a look-alike doesn't displace the correct, text-disambiguated answer, but no guard is perfect.
If a look-alike keeps landing at the top
If a visually-similar wrong product keeps getting promoted over the correct answer, contact MyChatBot support on Telegram (https://t.me/mychatbot_support). Include your account, the product integration id, the customer image, and the returned-vs-expected product ids so support can tighten relevance for your catalog. In the meantime, have the agent show the top image and ask the customer to confirm — that catches misfires before they reach a wrong answer.
Visual similarity tells you images differ, not why
Two genuinely different products (e.g. two similar dresses) can look very close to an image-matching system. Keeping deeper visual matching on helps where it matters, and confirming with the customer via the top image is the reliable safety net — especially since the match confidence isn't exposed as a readable score.
Turning on image search
Image search only works on catalogs that were indexed with it enabled. It's available across the common inventory integrations: WordPress/WooCommerce products, OpenCart products, XML/YML/text feeds, and spreadsheet (Google Sheets) inventories.
In the app: enable image indexing
When you connect (or edit) a product / inventory knowledge base, turn on the Allow image search toggle (shown as Index images on some connectors — "Include all images in your index"). This builds the image index; expect a longer initial processing run.
Knowledge bases live at https://app.mychatbot.app/knowledge-base (open a product feed to edit its settings).
To add image search to a catalog that was indexed without it, re-connect the feed with the toggle on. If a clean full rebuild is needed, delete and re-add the integration, or contact MyChatBot support with the integration id to force a full re-index.
Indexing constraints
- Up to 20 images per product are indexed.
- Image URLs must be publicly downloadable during indexing — unreachable URLs are silently skipped, thinning your index.
- For Google Sheets inventory, put image URLs in columns named
image1,image2,image3, and so on.
Using it: the agent and the Product MCP
Once a catalog is image-indexed, your agent's product search gains an optional image input — the agent can search by text, by image, or by both, depending on what the customer sends. Inside MyChatBot this is automatic.
If you're driving the same search from an external MCP client (Claude Code, Cursor, or any MCP host) via the Product MCP, the product-search tool accepts an image alongside — or instead of — a text query:
- Image only — pass the customer's image URL; the text query is optional.
- Image + text — pass both when the customer gives a picture and a description (e.g. a photo plus "red, under 3000"); this gives the best precision.
A strong image match comes back as the first product in the list, not as a separate score field. See Product MCP for the connection URL and the copy-paste setup command (also surfaced in the app on the Knowledge Base and Agents → Connectors screens).
Best practices
- Do keep deeper visual matching on for catalogs with many visually similar SKUs; don't leave it on for latency-sensitive flows where duplicate detection is all you need. Remember, changing it means a re-index.
- Do send text alongside the image when the customer provides both — combining the two beats either signal alone.
- Don't rely on image search for a catalog you haven't re-indexed with image search enabled.
- Don't build agent logic around reading a per-result "confidence" — the search returns a product list; use rank position (top = best match) plus a top-image confirmation instead.
- Do ask MyChatBot support to tighten relevance (with your integration id and the wrong-vs-right product ids) before disabling features if you see wrong matches — it's the least invasive fix.
- Do verify images are publicly reachable — unreachable URLs are dropped at index time.
Test it
- Turn on Allow image search for a test catalog and wait for indexing to finish.
- Take a known catalog product image, alter it slightly (light crop or a screenshot), and run a product search with that image.
- Confirm the correct product returns at or near the top of the list. (There's no confidence score, so rank position is your signal — a lightly-modified copy is trusted enough to be pushed to the top.)
- Repeat with a look-alike (a different product with a similar layout) and confirm it does not override a matching text query.