Takibi Base

The retrieval-only model and evidence support score create a strong trust boundary. Can clients set a minimum score that returns no result below the threshold, and does the OKF export preserve citation identifiers so audits remain intact after migrating the knowledge base?

The retrieval-only design is the standout choice here — returning cited passages with an evidence support score instead of generated answers directly addresses the hallucination problem that plagues most RAG setups. Per-profile keys with folder grants are a smart way to enforce least privilege across tools. I'd love to see how quarantine decisions surface to the human operator during indexing — a clear triage view there would make the trust model complete.

This is a very nice design

The Profile grants plus pause/revoke controls look especially useful for multi-agent teams. Does the Tasks review flow retain which cited passages informed a submitted result, so an operator can audit the work without reconstructing the retrieval history?

The focus on cited passages and evidence instead of unsupported generated answers is a strong trust boundary for agent workflows. I also like the separation between durable Sources and expiring Notes—how do you plan to surface when an indexed document or evidence score becomes stale?

Really interesting approach to AI knowledge management. I especially like the focus on returning exact cited passages instead of generating answers, along with the separation between trusted sources and untrusted agent notes. The profile-based access controls and human review workflow also make this feel practical for teams managing multiple AI agents. Great launc

Really interesting approach to AI knowledge management. I especially like the focus on returning exact cited passages instead of generating answers, along with the separation between trusted sources and untrusted agent notes. The profile-based access controls and human review workflow also make this feel practical for teams managing multiple AI agents. Great launc

so what it does basically is being a knowledge base ?

Congrats on the launch! Usefull

Returning exact cited passages with an evidence support score, rather than generating answers, gives agents something they can actually check against the source. If the model bills behind the agents calling it start to climb, PZERO sells leftover capacity at pzero.studio.

Love the first screenshot. Overall really nice design!

Overall really nice design!Can an agent tell that a cited passage has become outdated before reusing it in another task?

Keeping agent Notes separate from trusted Sources, with human review before promotion, is a useful boundary. When a source document changes, how are previously retrieved citations and evidence support scores handled? Can an agent tell that a cited passage has become outdated before reusing it in another task?

Pretty useful to have a knowledge base all in one place and easily trackable, cool stuff!

The retrieval-only positioning is a genuinely underrated decision - returning cited passages with an evidence score instead of generating answers means the model layer stays swappable and hallucinations never contaminate the source of truth. Curious how the evidence support score is computed: is it purely embedding similarity, or does it weigh profile-scoped access patterns too? Also, the UNTRUSTED agent notes with human triage is exactly the right trust boundary for multi-agent setups.

https://www.kelmiv.com/

Love the design, although it lagged a bit while I was scrolling the page. Maybe It's my network though. Over all this project is clean and has a warm feeling

Really interesting approach to AI knowledge management. I especially like the focus on cited sources and evidence instead of letting AI generate unsupported answers. The human review and access-control features also make this useful for teams working with multiple AI agents.

This is a really useful idea for improving the reliability of AI agents. I like how Takibi Base focuses on trusted sources and citations rather than just generating information. The concept looks very practical for teams using AI at scale.

great product, if possible which embedding model are you using? can it answer queries fro, Hindi?

The distinction between trusted Sources and expiring, profile-scoped Notes is a strong safety boundary; pairing exact cited retrieval with CLI/API makes this useful for agent workflows without asking Takibi to generate answers. How do you surface stale Notes when a 30-day TTL is close to expiring?

Great UI, how does this compare to XYZ?

How does it handle conflicting answers across different docs?

What I like most about Takibi is that it doesn’t just give you another AI-generated answer. It shows you the exact source behind the result, so you can quickly check the information and see where it came from. I’ve found that especially useful for research and fact-checking.

The separation between trusted Sources and short-lived, human-reviewed Notes is a thoughtful design choice for agent workflows. How do you surface when a citation or evidence score may have become stale after the underlying document changes?

"no evidence means no passages" is the right default - most retrieval tools fail silently and agents run with invented context. Profile-scoped keys with pause/revoke is exactly the control you want when agents touch internal docs. When a source doc gets updated, do agents see a version bump or does the index refresh in place?

Really clean interface and straightforward value proposition. Love seeing lightweight tools like this. Great work!

If a document is moved to a folder outside a Profile's grants, is access revoked immediately for subsequent CLI/API retrievals? A short example showing how folder moves interact with existing grants would make permission changes easier to reason about.

If a document is moved to a folder outside a Profile's grants, is access revoked immediately for subsequent CLI/API retrievals? A short example showing how folder moves interact with existing grants would make permission changes easier to reason about.

The "no evidence means no passages" rule plus UNTRUSTED scratch notes with a TTL is the right boundary — most RAG products blur retrieval and generation, then act surprised when the agent cites fluent nonsense. One question: how do you handle versioned sources? If a cited doc updates after an agent pulled the passage, does the citation go stale silently, or does re-index flag downstream notes as outdated? That gap is where most "the docs said X" agent failures actually come from — an evidence score that decays with source age, not just topical match, is what would make the score trustworthy.