Hi Ashish,
Thank you for the detailed information. Your general setup is correct, but the PDF has not yet been successfully added to the searchable context.
The PDF workflow has two separate stages:
- Upload and text extraction: AxiaChat reads the PDF and creates temporary text chunks. The “47 chunks” message confirms that this stage completed.
- Context indexing: AxiaChat generates embeddings and stores the chunks in the selected context. The Context statistics are calculated from this final index.
Therefore, 0 documents / 0 chunks / 0 tokens means that the second stage did not complete successfully. A progress value of 100% only means that the indexing queue finished; it does not necessarily mean that embeddings were created.
Recommended workflow
- Go to Training → Context.
- Select the USA bot.
- Select USA Physician Pathways as its active context.
- Keep Storage set to Local.
- Upload the PDF in the PDF/TXT source section.
- Wait until the file row shows 47 chunks, not only 100% upload progress.
- Without refreshing the page or changing the bot/context, click Save & Index.
- Wait for all indexing batches to complete.
- Refresh the document list and Context statistics.
No separate “attach” operation should normally be required: a PDF uploaded on that screen is automatically included when Save & Index is clicked.
There is currently an important edge case: if the page is refreshed or the selected bot/context is changed after uploading but before indexing, the uploaded PDF may no longer be present in the current browser selection. In that case, upload the PDF again and click Save & Index immediately after chunking completes.
If indexing still finishes with zero items, please run the Embedding diagnostic under Settings → Advanced → Diagnostics. This usually indicates an embedding-provider problem, such as:
- Missing or invalid API key.
- Provider quota or billing restriction.
- An embedding API request being blocked by the server.
- Claude being used without the OpenAI fallback required for embeddings.
For verification, the Context document list should show the PDF, normally grouped as one file, with a non-zero chunk count. You can use View to inspect the extracted chunk text. Then open the USA bot preview and ask a question whose answer appears only in that PDF. Keeping “Include Current Page Content” disabled during this test helps ensure that the answer is coming from the PDF context.
Your other settings are suitable for a static PDF knowledge base:
- Local storage: correct.
- Knowledge Sources Limit 5: a reasonable default.
- Include Current Page Content disabled: correct for an isolated test.
- AutoSync disabled: correct for manually uploaded PDFs.
- Context Max Length 4096: valid, although newer releases use a higher default and approximately 6000 characters may provide richer context.
For the country knowledge bases, we recommend a separate context for each country because physician pathways and regulations can overlap semantically while having materially different answers. Separate bots are also recommended when each country needs different instructions, branding, deployment, or conversation behavior.
A bot can use one active context at a time. If you combine all countries into one context, every document should clearly identify its country, but retrieval isolation will be weaker. For this use case, a structure such as the following is safer:
- USA bot → USA Physician Pathways context
- Canada bot → Canada Physician Pathways context
- Australia bot → Australia Physician Pathways context
- Netherlands bot → Netherlands Physician Pathways context
- Singapore bot → Singapore Physician Pathways context
- General Finance bot → General Finance context
Before creating the remaining contexts, please confirm that the USA Context shows non-zero indexed chunks and that the embedding diagnostic passes. If it still returns zero, please send us the AxiaChat version, the indexing batch log, and the Diagnostics ZIP through the Pro support channel. Please do not post the PDF or any sensitive data publicly.
Best regards,
