How does Takibi Base return cited passages to AI agents?

Takibi Base answers an agent's question by searching the documents you've uploaded and returning the matching passages themselves — not a rewritten summary. Each passage comes with a citation (document ID, document name, chunk ID, and title) plus an evidence support score. If it finds no matching evidence, it returns no passages instead of guessing. This applies to questions asked through the CLI or API against documents in your own workspace.

What the agent actually receives

When a tool asks a question, Takibi searches the sources that the Profile's keys are allowed to read. The response is structured, so the calling agent can quote the source text and point back to where it came from.

From the documented CLI example:

$ takibi ask -q "What is our refund policy?" --json
{
  "spans": [{
    "text": "Customers may request a refund within 30 days of purchase.",
    "documentId": "00000000-0000-4000-8000-000000000010",
    "documentName": "refunds.md",
    "chunkId": "00000000-0000-4000-8000-000000000090",
    "title": "Eligibility"
  }],
  "citations": [{ ... }],
  "support": 0.4,
  "supportVersion": "v1",
  "abstained": false,
  "candidateCount": 1,
  "answerability": "unknown",
  "conflict": false
}

The key fields:

The evidence support score is described as "a measure of support in the retrieved evidence." In the sample, a single cited passage produced a support value of 0.40 — so a low score with one candidate is a signal to check the source rather than trust the answer outright.

Why passages instead of generated answers

Takibi's stated design is "The source. In its own words." The agent gets the source text and a citation for each passage, which means the grounding happens at the retrieval layer: the tool decides what to do with the passage, but it can always show where the text came from. This is what makes answers citable rather than paraphrased.

The flip side is the abstention behavior: if no matching evidence is found, Takibi returns no passages. There is no fallback to a generated guess, so an empty result is a meaningful signal that the documents don't cover the question.

What has to be in place first

Retrieval only works over documents that are indexed and permitted:

1. Upload documents into projects and folders. Takibi converts them to Markdown so agents can search them. You can download the originals or extract their text again.
2. Check indexing status. Files show as indexed, converting, or quarantined. Only indexed content is searchable.
3. Grant access through a Profile. A Profile holds a tool's keys and document permissions. Open the Profile's Access tab, choose Edit, and select the folders it can read. Every key in that Profile uses those permissions.
4. Account for future content. Folder access includes documents you add later, once they are indexed. Project access also includes future folders and documents outside folders — Takibi asks you to confirm this broader access.

Separate Profiles are the intended way to give different tools different access, so a tool that only needs refund policy documents doesn't search everything.

Common points to watch

• Uploads, imports, downloads, and billing require your own workspace. The sample workspace is for trying things out, and edits reset on refresh — so demo results won't persist.
• A folder grant is not retroactive to unindexed files. New documents become readable only after indexing completes.
• A low support score with few candidates is worth inspecting. The response exposes candidateCount and conflict precisely so the calling agent can decide whether to rely on the passage.
• Project-level access is broader than it looks. It reaches future folders and documents outside folders, which is why Takibi requires explicit confirmation.

To see the retrieval behavior before wiring anything up, the sample workspace lets you ask questions and change access, with the caveat that changes reset on refresh.