Changelog
What changed
Last updated
Changes a customer would notice, newest first. Anything that changes what the API returns is announced here before it ships, with the old shape kept working.
Add someone from a transcript you already have
- When you add someone, choose "From a transcript you already have": paste it, or pick a .txt, .vtt or .srt file (a Meet, Zoom or Teams transcript, a Granola or Otter export). Keeplore shows the speakers it found, you say which one is the person, and confirm they were told.
- Only that person's words become rules. What everyone else in the conversation says stays as context, shown next to each rule it preceded.
- The person gets a link to the rules written under their name and can flag any that isn't right, the same as after an interview. Keeplore keeps the words, never a recording.
- The interview page is redone: the conversation reads as a transcript, no bubbles, and the rules Keeplore takes from your words appear beside it as you answer, each with the quote behind it. On a phone they sit behind a count above the answer box.
- Speak your answers: a Speak button in the interview uses your browser's own speech recognition (Chrome, Edge, Safari), your words appear in the box to edit before sending, and questions can be read aloud. Nine languages.
- A manager whose own interview is open gets a Continue your interview button on the dashboard.
- A sample pilot scorecard at /scorecard/sample, for Acme, so a buyer sees the result page before paying for the pilot.
- New API keys start with kl_. Keys made before today start with lore_ and keep working.
- Fixed: since the evening of 27 October, an interview link for someone added with their email address, including the manager's own 10-minute interview, opened to "Something went wrong". Interview links no longer touch sign-in at all.
Scoped keys, an MCP server, and a record of who approved what
- Every rule is now internal or customer-safe. Internal is the default, so a rule that names a colleague never reaches an agent that talks to customers unless a manager marks it customer-safe on purpose.
- API keys can be limited to customer-safe rules, and can leave people's names out of answers and exports.
- An MCP server at POST /api/mcp: Claude Desktop, Claude Code, Cursor and other MCP clients connect with one URL and the same key. Tools: ask_keeplore, search_rules, list_topics.
- POST /api/v1/query returns handoff (true when no approved rule applies) and found (how many rules were considered). Error responses carry a stable code.
- Every approve, reject, edit, confirm and retire is recorded with who did it. Edits keep the original wording, and edited rules are marked in the API and the export.
- Rule upkeep: the Knowledge page lists rules nobody has confirmed in six months, and rules that name someone who has left. Confirm or retire them in a click.
- When you add someone, you can set what the interviewer asks about first: questions your agents couldn't answer, a case that was escalated, a scanner finding.
- A weekly digest on Mondays to every admin: rules approved, waiting for review, agent questions and the share answered, rules that need a look.
- The dashboard opens with the numbers that matter and a getting-started list. The Knowledge page has search, an audience filter and downloads. A Team page for inviting colleagues. A Send a reminder button on open interviews.
- A health check at GET /api/health for uptime monitors. It also says whether the database has every migration the code expects.
- POST /api/v1/feedback and the report_outcome MCP tool: your agent tells Keeplore when a person followed or overrode a rule. Three overrides in 30 days list the rule for a look, with the notes. Nothing is watched.
- The written baseline: set your public help centre on the Integrations page and Keeplore keeps its articles next to your rules. A new rule that contradicts an article is flagged for the reviewer with the article's words, and query answers carry matching documents and a grounded field (rules, documents or none).
- Four questions on the first visit (which agent, where, why now, how many interviews) that tailor what the dashboard suggests first. Skippable.
- An interview test set (npm run eval:interview) runs the real prompts against fixed transcripts on every prompt change. Its first run caught extraction writing a Dutch interview's rules in English; fixed.
- The Pilot Scorecard at /app/scorecard: your 30 questions (20 in scope, 10 controls), frozen and dated; your agent's answers pasted before and after the pilot, with Keeplore's own recorded alongside; blind grading by your team lead on four outcomes with wrong-and-confident counted on its own; the result in counts on a private page to forward. Once a month Keeplore re-answers the questions and the digest says what changed.
Follow-through on every interview
- The person interviewed sees every rule written under their name after Finish, and can flag one as wrong with a note. Flagged rules go back to review.
- A manager summary when an interview finishes, written to be forwarded. One reminder to the interviewee after two quiet days, with the same link.
- An Agent questions page: what agents asked in the last 30 days and what Keeplore couldn't answer, most asked first.
- Approve all skips rules that contradict an approved rule. Only admins read full transcripts. Agent questions are deleted after a year.
- Ask what your agent would ask, from the Knowledge page, without an API key.
An interviewer that listens
- One question per message, a straight answer about how long is left, and no asking for a real case three times.
- Rules are labelled practice, plan or guess. Guesses and quotes that are only "yes" or "sure" are dropped; plans are capped at 50% confidence.
- Approve all only takes rules Keeplore is at least 60% sure of. Every rule shows the question its quote was answering.
- A 10-minute self-interview for the manager who sets Keeplore up.
Keeplore launches
- Interviews, review and the query API. Two free interviews with a work email. The Gap Scanner at /scan.
Related
Try it before you decide.
Pick one person whose knowledge your agent is missing. Send them the interview link. Read the rules that come back.