> For the complete documentation index, see [llms.txt](https://boundaryai.gitbook.io/boundaryai-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://boundaryai.gitbook.io/boundaryai-docs/bringing-in-your-feedback/uploading-an-existing-dataset.md). # Uploading an Existing Dataset If you have already collected feedback outside BAI Analytics (survey exports, support-ticket dumps, interview transcripts, or document libraries) you can upload it directly into a Feedback Group. The platform parses your file, infers what each column or section means, runs the same analysis as native surveys, and lands the result in your dashboards within minutes. This page covers the upload flow end-to-end: supported formats, column mapping, document ingest, and what happens after import. For other ways to bring feedback in, see [Creating a Survey](/boundaryai-docs/bringing-in-your-feedback/surveys/editor.md), [Connecting a platform](/boundaryai-docs/bringing-in-your-feedback/connect-to-your-existing-systems.md), and [Web scraping](/boundaryai-docs/bringing-in-your-feedback/social-listening.md). For the container that holds uploaded data, see [Feedback Groups](/boundaryai-docs/feedback-groups/feedback-groups.md). *** ### When to upload Use the upload flow when: * You have **historical data** in spreadsheets you want to analyse alongside new feedback. * You have **interview transcripts** (text or JSON) and want them clustered into themes. * You have **a stack of documents** (PDFs, Word files) that contain qualitative feedback you would otherwise have to read manually. * You have **call recordings or voice memos** and want them transcribed and analysed automatically. * You exported responses from another survey tool and want to keep tracking them in BAI. Uploads land inside a Feedback Group as a source, so they participate fully in cross-source analytics, Grouped Themes (previously called Super-Themes), Custom Monitoring, and reports. *** ### Supported file types BAI Analytics accepts three broad classes of files: #### Structured files (column-mapped) | Format | Extension | Notes | | ---------------------- | --------------- | ---------------------------------------------------------------- | | Microsoft Excel | `.xlsx`, `.xls` | Multi-sheet workbooks supported; you pick which sheet to import. | | Comma-separated values | `.csv` | UTF-8 recommended; other encodings auto-detected. | Structured files go through the **mapping flow**: BAI Analytics shows you a preview of your columns and suggests a question type for each, which you can review and adjust before importing. #### Document files (auto-extracted) | Format | Extension | Notes | | -------------------------- | --------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | PDF | `.pdf` | Text-based PDFs; scanned/image-only PDFs are not OCR'd. Speaker detection supports transcripts. | | Word documents | `.docx` | Tables and lists preserved. Speaker detection supports transcripts. | | JSON interview transcripts | `.json` | One conversation or thousands per file; the shape is detected automatically. Speaker roles are read from the file; supports common call-centre and meeting-tool exports. | Document files skip column mapping and go through an AI conversion step that extracts question-and-answer pairs (or theme/evidence pairs for free-form transcripts) before being analysed like any other upload. #### Audio (auto-transcribed) Upload a call recording or voice memo (`.mp3`, `.wav`, `.m4a`, `.aac`, `.flac`, `.ogg`, `.opus`, `.amr`, `.caf`) through the same upload dialog as any other file. BAI Analytics transcribes it automatically, with speakers separated, then runs the transcript through the same analysis as a document upload. Long recordings are chunked automatically, so there's no need to split a call in advance. Upload one audio file at a time. #### Limits * **File size**: up to **500 MB** per file. Large files are uploaded directly to cloud storage in chunks, so a slow network does not force a restart. * **Columns**: up to **1,000 columns** per spreadsheet. Beyond that, the mapping table becomes unwieldy in the browser. * **Rows**: no hard cap; very large files just take longer to process. * **Multi-document upload**: drop multiple `.pdf` or `.docx` files at once and BAI Analytics merges them into a single survey with a *Source* column tracking which file each row came from. Audio files are uploaded one at a time. *** ### The upload flow The flow has up to four steps, depending on what you upload. Single-sheet files skip step 2; documents skip steps 2 and 3. #### Step 1: Upload your file 1. Open the Feedback Group you want the data to land in. 2. Click **Upload Data**. 3. Drag-and-drop your file into the upload area, or click to pick from your device. Files start uploading immediately. Large files use a resumable upload session, so the progress bar reflects real bytes-on-the-wire rather than waiting for the whole file to finish. #### Step 2: Pick a sheet *(multi-sheet Excel only)* If your workbook contains more than one sheet, BAI Analytics shows a sheet picker. Select the sheet that holds your feedback data and continue. Single-sheet workbooks skip this step automatically. #### Step 3: Review the column mapping BAI Analytics scans the first rows of your file and **auto-suggests a field type for every column**, using both the column header and a sample of the values. Each column gets a dropdown so you can confirm or override the suggestion. See Column mapping in detail below for the full list of field types and how the AI inference works. #### Step 4: Confirm survey details Fill in: * **Survey name**: required. Used to identify this dataset inside the Feedback Group; pick something descriptive (e.g. *"Q1 2026, Customer Satisfaction Export"*). * **Survey language**: required. Sets the source language for sentiment, theming, and translation. Affects analysis quality, so set it accurately. * **Selected file**: read-only confirmation of which file is about to be imported. If the destination Feedback Group has monitors with **Auto-cover new sources** switched on, this step also lists which of them will watch the uploaded data. There is nothing to configure: coverage is automatic, and detection runs as part of the first analysis. To change what's monitored, use the group's **Manage monitoring** panel; see [Custom Monitoring](/boundaryai-docs/analysing-your-feedback/custom-monitoring.md). {% hint style="info" %} **Where did the monitoring step go?** The upload wizard used to end with a monitor-selection step. Monitoring now lives at the Feedback Group level and covers new uploads automatically, so that step no longer exists. {% endhint %} *** ### Column mapping in detail This is where the platform learns what your spreadsheet actually contains. #### Available field types | Field type | Use it for | | ------------------- | -------------------------------------------------------------------------------------------------------------------------------------- | | **Single Choice** | Columns where every cell is one of a small set of values (Yes/No, satisfaction labels). | | **Multiple Choice** | Columns where each cell may contain several values, often comma- or pipe-separated. | | **Short Answer** | Brief free-text: names, one-line comments, identifiers. | | **Long Answer** | Detailed open-ended responses (the columns the AI will analyse for themes and sentiment). | | **Linear Scale** | Numeric ratings on a custom range (1–5, 1–7, 1–10 outside NPS). | | **NPS** | The standard 0–10 likelihood-to-recommend column. | | **Metadata** | Context columns that should not be analysed but should be available for segmentation: IDs, timestamps, demographics, channel, country. | | **Ignored** | Columns that shouldn't be imported at all. | The richest analytical signal comes from **Long Answer** columns; metadata columns power segmentation in the analysis view. #### How auto-mapping works The auto-mapping is driven by AI, backed up by a few reliable rules: * **AI inference** reads the column header and a sample of the values for context, then proposes a type. * **Pattern recognition** catches clear cases the AI sometimes misses, for example, columns whose values are only *Yes / No / True / False* are treated as **Single Choice**, and headers like *"Reason for…"*, *"Tell us about…"*, or *"Comments"* with sentence-length answers are treated as **Long Answer**. * **Multilingual support**: BAI Analytics recognises common patterns in English, French, and Spanish, so a column called *"Pourquoi recommanderiez-vous…"* is auto-detected as a long-answer free-text column without needing manual review. * **Conservative default**: columns the AI cannot confidently classify default to **Metadata**, which is non-destructive: they will not be analysed, but their values are still imported and remain available for segmentation. #### Required fields Before you can finalise the mapping, BAI Analytics checks that: * The survey has a **name**. * At least **one column is mapped to a question type** (i.e. not all columns are Metadata or Ignored). * The file contains **at least one row of data**. If any of these fail, the validation dialog tells you exactly which columns to revisit. You can fix the mapping in place and continue without re-uploading the file. #### Tips for clean mapping * Use **clear, human-readable column headers** before uploading. *"Q1\_LongAns"* gives the AI nothing to work with; *"Why did you choose this option?"* is unambiguous. * **Split combined fields** into separate columns. A single *"Major, Semester"* column should be two columns (*Major* and *Semester*) so each can be mapped independently and segmented separately. * **Use standardised date formats** (`YYYY-MM-DD` or your local long format). Excel serial-number dates are detected and converted, but consistent formatting reduces edge cases. * **Mark internal columns as Ignored** rather than deleting them. That way you keep one master file and the upload still imports cleanly. * **Identifiers stay consistent across uploads** if you intend to update or join data later. *** ### Document and transcript ingest When you upload a `.pdf`, `.docx`, or `.json` interview file, BAI Analytics bypasses column mapping and runs an AI conversion step instead. #### Documents (PDF, DOCX) The system extracts text from the document and detects speakers in transcript-style files so you can assign roles. Where the file is structured as question-and-answer pairs, BAI Analytics preserves that structure during import. The output then runs through the same theme-and-sentiment analysis as a normal upload. PDFs must be **text-based**. Image-only or scanned PDFs are not currently OCR'd; convert them to searchable PDFs first. #### Multi-document upload Drop multiple `.pdf` or `.docx` files at the same time and BAI Analytics processes them in parallel, then merges the results into a single survey. A *Source* column on every imported row tracks which document the text came from, so you can segment the analysis by document. This is useful for things like: * A library of customer interview write-ups (one Word doc per interview). * A folder of feedback emails or letters exported as PDFs. * A set of focus-group transcripts. #### JSON conversation exports For transcripts produced by call-centre tooling or meeting platforms, BAI Analytics accepts JSON exports, and a single file can hold **one conversation or thousands**. On upload, the file's shape is detected automatically: a single conversation, an array of conversations, or a dictionary of conversations keyed by ID. Speakers are detected and their labels preserved, so the analysis can distinguish what the participant said from what the interviewer said. Very large transcripts are supported, and consecutive turns from the same speaker are merged so the AI sees full statements rather than fragmented utterances. Two edge cases are rejected with a clear message rather than imported badly: **tabular JSON** (plain records rather than conversations) should be converted to Excel or CSV and uploaded through the mapping flow, and a **mixed bundle** of JSON plus Word/PDF files should be split into two uploads. #### Choosing how conversations become rows When an upload contains two or more conversations or documents, the details step offers a choice of **processing mode**: * **Split into passages**: the AI breaks each conversation into thematic passages, one row per passage. Best for depth on a handful of long conversations; available for bundles of up to 50. * **One row per conversation**: each conversation or document becomes a single response, summarised during analysis. Built for scale across hundreds or thousands of conversations. BAI Analytics recommends a mode based on the size of the upload (around ten conversations is the tipping point), but the choice is always yours. Single-conversation and single-document uploads skip the choice entirely; nothing changes in those flows. #### Joining a metadata sheet Upload an Excel or CSV **lookup sheet together with a JSON export** and BAI Analytics joins the sheet's columns onto the conversations as metadata. Pick the join key on both sides (an ID, phone number, or email that appears in both files); matching is exact, ignoring case and surrounding whitespace. The joined columns land as segmentation metadata on each conversation, and a column whose name collides with an existing field gets a *"(sheet)"* suffix. This is how you attach CRM context (region, plan, agent, satisfaction score) to raw conversation exports in one pass, with no manual re-keying. *** ### What happens after import Importing is the start, not the finish. Once the file is in, BAI Analytics runs an automatic process that: 1. **Cleans the text**: removes stray control characters, normalises encoding, and strips obvious noise. 2. **Detects the language** of each response (which can differ from the survey's source language for mixed datasets). 3. **Scores sentiment** for every open-ended answer. 4. **Detects themes** by clustering similar answers and labelling each cluster. 5. **Runs Custom Monitoring** for every group monitor that covers the new source (see [Custom Monitoring](/boundaryai-docs/analysing-your-feedback/custom-monitoring.md)). 6. **Enriches metadata**: source provenance, timestamps where inferable. A progress toast in the corner of the screen tracks the analysis through each stage. You can keep working while it runs; when processing completes, the survey appears in the Feedback Group's source list and the *Open analysis* link becomes active. If the analysis fails part-way (rare: usually a malformed cell or a network blip), the toast reports the error and you can re-upload the file once the underlying issue is fixed. *** ### Validation, errors, and recovery #### Errors that block import These stop the import outright; you must fix them before continuing: * **No question column**: all columns mapped to Metadata or Ignored. Map at least one column to a question type. * **Empty file**: no rows of data after the header. * **Bad encoding**: the file is not in a text encoding the parser recognises. Re-export as UTF-8. * **Corrupted or password-protected file**: re-export without the password. * **File over 500 MB**: split into smaller files or remove unused columns. #### Warnings (non-blocking) These let you continue but call out potential issues: * **Sparse rows**: many empty cells. The import proceeds; affected rows simply contribute less signal. * **All-empty column**: defaults to Metadata; you can change it before continuing. #### Recovery If the validation dialog catches an issue, you can **fix the mapping in place and retry without re-uploading**. The file stays parsed in memory between attempts, so corrections are quick. *** ### Best practices * **Clean column names before uploading.** *"How likely are you to recommend us?"* gives the AI a much stronger signal than *"Q4\_NPS"*. The five minutes you spend tidying headers saves you ten minutes of mapping review. * **Keep one column = one variable.** Compound columns like *"Major, Semester"* should always be split into separate columns. * **Standardise dates and identifiers.** A consistent ISO date and a stable user/email column keep cross-upload joins clean. * **Mark unused columns as Ignored, don't delete them.** Keeping the master file intact makes future re-uploads predictable. * **Set the survey language correctly.** Mismatched language degrades sentiment and theme quality more than people expect. * **Use multi-document upload for qualitative libraries.** Don't manually concatenate ten interview docs into one, drop them all in and let BAI Analytics keep the source attribution. * **Set up group monitors before big uploads.** Monitors with auto-coverage on watch a new upload from its first analysis pass, so the signals you care about surface immediately, with no re-run. * **Match the upload to the right Feedback Group.** Uploading into *Individual Surveys* is fine for a one-off, but a recurring dataset belongs in a custom group so it can be tracked over time.