OCR for Your Chatbot Knowledge Base: Getting Text Out of Scans, Photos and Menus

Many small businesses keep their most useful information in the least searchable form: a photographed price list, a scanned brochure, a menu saved as an image, a PDF exported from a design tool with the text outlined. To a person these are perfectly readable. To an AI assistant they are blank: a picture of text is not text. Unless something converts the image into words, the assistant cannot answer questions about it, and it may confidently answer from older documents instead. This guide explains how text is extracted from images and scans, the two main approaches, optical character recognition (OCR) and vision AI models, how to choose between them, how to check the quality of what comes out, how to handle languages and tables and how to prepare files so the result is good. It also gives you a simple rule: whenever you can, provide the text itself.

Why image-only files are a problem

The first test is simple: open the file and try to select and copy a sentence. If you cannot, it is an image.

Approach 1: classic OCR

OCR software analyses an image, finds lines and characters and outputs text with a confidence score. Open-source engines such as Tesseract are widely used, work offline and cost nothing per page. They are good on clean, printed text in common languages, and weaker on handwriting, unusual fonts, low-resolution photos, curved pages and complex layouts.

Approach 2: vision AI models

Multimodal models can look at an image and describe it: transcribe the text, read a table, explain a diagram and note what a logo or badge means. They handle photos, layouts and messy pages better than classic OCR, and they can add useful context. The costs are per-image usage charges, some latency and a privacy consideration, since the image is sent to a model provider. With a bring-your-own-key setup, you choose the provider and pay them directly.

How Crescelo handles images and scans

Crescelo can read text from images and scanned documents when they are added to the knowledge base. It uses OCR for text extraction and, for images, can use a vision model through your connected provider to transcribe and describe the content. Image OCR and multimodal ingestion are part of the Growth plan. OCR reads English by default, so for other languages check the current language settings or provide text versions. Provider usage for vision is billed by your provider under your key.

Choosing between them

Always check what came out

1. Open the extracted text next to the original.
2. Check every price, number, date and phone number; these are where mistakes hurt.
3. Look for missing sections, merged columns and scrambled table rows.
4. Check names and unusual words.
5. Fix errors in the source text or re-scan.
6. Ask the assistant the questions the document answers and compare.

OCR confidence scores help: low-confidence pages deserve a closer look. But a high confidence score does not guarantee that a price is right, so check the numbers yourself.

Prepare files to get good results

• Scan at a reasonable resolution, such as 200 to 300 dots per inch for documents.
• Keep pages straight, evenly lit and free of shadows.
• For photos, shoot straight on, close enough to read, in good light.
• Avoid compression that blurs small text.
• Split multi-page scans into a clear order.
• Prefer the original digital file if you have it: export a PDF with real text, not a flattened image.

Better still, retype the key facts. A one-page text document of your prices is more reliable than any OCR of a photo.

Tables and layouts

Tables are the hardest case. OCR can flatten them into a stream of numbers in reading order, and the association between a row label and its price can be lost. Vision models do better, but check every row. For important price lists, convert them to plain text sentences such as 'Colour treatment, medium length: $95, about 2 hours'. See document chunking strategies and preparing your service menu and pricing documents.

Languages

• OCR needs the right language pack and script.
• Mixed-language documents can produce errors.
• Right-to-left and non-Latin scripts need engines that support them.
• Accented characters must be preserved, since they change words and names.

If you serve customers in several languages, test each. See multilingual chatbots.

Privacy and cost

• Scans may contain personal or confidential information: remove or redact what the assistant does not need.
• Decide whether images may be sent to a provider, and check its terms.
• Vision models cost per image, so process only what you need.
• Delete source images after extraction if you do not need them.

Worked example: a restaurant menu

A restaurant's menu existed only as a photo of a chalkboard and a design-tool PDF with outlined text. Uploaded as they were, the assistant knew nothing about prices or allergens. The owner photographed the board straight on in good light, used a vision model to transcribe it, and checked every price against the board. She also retyped the allergen notes as text. Now the assistant answers 'is the salmon gluten free?' and 'how much is the tasting menu?' correctly, and she keeps a plain text version for future updates. The photo took a minute; the checking took twenty.

Frequently asked questions

See the knowledge base feature, try uploading a document in the playground and read how to add your knowledge to a chatbot.

Further reading

• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al.): The original research paper behind retrieval-augmented generation.
• Google Search Central: creating helpful, people-first content: What Google looks for in useful pages.