diderot
Health Gecti
- License — License: Apache-2.0
- Description — Repository has a description
- Active repo — Last push 0 days ago
- Community trust — 40 GitHub stars
Code Basarisiz
- rm -rf — Recursive force deletion command in .goreleaser.yaml
Permissions Gecti
- Permissions — No dangerous permissions requested
Bu listing icin henuz AI raporu yok.
Search and ask across your organization's knowledge: Slack, coding-agent sessions, docs and tickets. Self-hosted.
Diderot
Diderot answers questions about what your organization has written down. It
ingests Slack, Discord, coding-agent sessions, Notion, Linear, Zendesk, and
folders of Markdown files into one Postgres event log, embeds everything into Qdrant, and runs an agent that
searches that corpus and answers with citations: in Slack, and as an MCP
server for your own agents and tools. You run one server, and administer it
in the browser at /admin.
Retrieval is vector-only: the question is embedded, Qdrant returns the nearest
chunks the caller is allowed to see, and results are deduplicated per event.
There is no reranker and no keyword index.
Features
- Sources. Slack, with live Socket Mode events and a continuous sync that
catches anything missed during restarts; image attachments are captioned and
PDF, DOCX, and XLSX attachments are extracted to text. Coding-agent sessions
(Claude Code, Codex, ATIF, Hugging Face STS), pushed over HTTP or polled from
a session feed, summarized and expandable to the exact records behind a hit.
Discord channels, threads and forum posts, over the gateway with a daily
re-read. Notion page trees, Linear issues and comments, Zendesk tickets,
comments, and Help Center articles, and the Markdown files in a mounted
folder. A new
source is one Go package and one registration line
(Adding a source). - Permissions. Every event carries a scope. Private Slack channels are
visible only to their members, sessions can be owner-only, and a policy file
grants groups access to restricted content. Every search is filtered to the
caller's scopes. - Ask. An agent that calls
retrieve,list, andexpandover the
corpus and answers with inline citations. Mention@diderotin a Slack
channel (each thread is one conversation) or DM it, or use the MCPask
tool or the HTTP API. Optional tools: web search (Tavily), code search
(Sourcebot), and the read-only tools of any MCP servers you configure
(deployment history, say), on Ask itself or on a delegated observability
investigator. - MCP server.
retrieve,list,expand,attachment, andaskfor
your own agents at/mcp, on every install, with the same bearer token the
HTTP API takes. - Slack extras. Thread-summary message shortcuts, scheduled reports
delivered to Slack, and optional investigations of PagerDuty alerts posted in
the alert's thread. - Any model provider. Add OpenAI, Anthropic, or any OpenAI-compatible
endpoint (a gateway, a local vLLM or Ollama) in the admin UI; Diderot
lists its models, and each job (embeddings, Ask, summaries, captions,
document extraction) gets the one you choose, tested before it is saved. - Authentication. OIDC bearer tokens from your IdP, static tokens, or a
loopback-only no-auth mode for local use. The browser UI signs in through
your IdP, sits behind an authenticating proxy, or signs in locally with a
static token, and a fresh install has a bootstrap admin for first setup. - Admin UI. Everything an administrator does is at
/admin: setup,
connectors and their channels, models, prompts, scheduled reports,
operations, and every Ask session. A separate page lets people choose
which of their agent sessions are private.
Quickstart
docs/quickstart.md runs Diderot on one machine with
Docker Compose: Postgres, Qdrant, a local embedding model, and the server.
You need Docker and an OpenAI or Anthropic API key; Ask sends each question,
and the corpus text it reads to answer it, to that provider. To ingest your
own data, add a Linear API key, a Notion integration, a Slack app, or a
folder of Markdown files; out of the box it can index Diderot's own
documentation.
Documentation
- Quickstart: a working install with Docker Compose.
- The
diderotcommand: connecting coding agents to a
server, and asking from the shell. - Setup: from the first sign-in to a live install.
- Running in production: what a deployment for a team
needs, how replicas and leadership behave, and a Kubernetes example. - Upgrading: moving an existing install to a newer
release. - Design: how Diderot works, as built.
- Configuration reference: every environment
variable. - Authentication and authorization: identity, scopes, the
policy file, and each surface's authenticator. - Slack app setup: creating the Slack app from its
manifest, connecting it, and what each permission is for. - Discord: creating the Discord bot, inviting it, and
choosing the channels to index and who may read them. - Agent-session ingest: accepted formats, the push
protocol, and the session feed. - Tool connections: giving Ask the tools of your MCP
servers, such as Grafana and Sentry. - Prompts: where the model instructions live and how to
change them. - Adding a source: writing a source of your own,
with the Markdown folder source as the worked example. - Contributing and AGENTS.md, for people and
coding agents working on Diderot.
Install, use, administer
Install by running the server: the quickstart with
Docker Compose, or in production from the container
image.
Use it from Slack (@diderot, or a DM) and over MCP. On your own
machine, the diderot command connects your coding agents
(Claude Code, Codex, Cursor, OpenCode, Pi) to a server, installs the agent
skill, and asks questions from the shell:
diderot setup https://diderot.example.com --harness claude-code
diderot ask "What did we decide about the billing migration?"
Any other MCP client points at https://<your-diderot>/mcp with a bearer
token the server accepts (step 5 of the
quickstart).
Its tools are retrieve (semantic search), list (enumeration by source,
person, Slack channel, and time), expand (the fuller source behind one
hit, and its attachments), attachment (one image or document), and ask
(a cited answer; pass its session_id back to ask a follow-up). The same
token works on the plain HTTP API: POST /v1/ask, /v1/retrieve,/v1/list, and /v1/expand-chunk.
skills/diderot is the agent skill that teaches a
coding agent when and how to use those tools; diderot setup installs it.
Administer it in the browser at /admin, the one place an install is
set up and run from.
What the diderot binary does
The container image holds one binary, diderot: the server, its
maintenance commands, and the client commands setup, doctor, ask, andsessions.
Releases carry a small diderot for people's
machines with the client commands alone, which reach a server over MCP and
can ship the machine's coding-agent sessions to it
(The diderot command). The server commands below run where
the server runs (docker compose exec diderot diderot ...,kubectl exec deploy/diderot -- diderot ...) because they read the server's
own configuration.
| Command | What it does | When to run it |
|---|---|---|
diderot serve |
Runs the server: the API, /mcp, the admin UI, and every background worker. It applies pending database migrations as it starts. |
Always; it is the container's command. |
diderot probe <healthz|readyz> |
Asks the local server whether it is up, or ready, and exits non-zero if not. | As a container health check or a Kubernetes exec probe (production). |
diderot qdrant-audit |
Compares the vector index with the database and queues what is missing or wrong for the server's workers to repair. | After changing the embedding model, restoring a backup, or losing the Qdrant volume. |
diderot reset-admin |
Turns the bootstrap admin sign-in (admin / admin) back on. |
When no administrator can sign in any more. Turn it off again from the setup checklist. |
Two more commands, ui-dev and session-dev, serve parts of the UI from
fixtures for people working on Diderot itself
(CONTRIBUTING.md).
Status
Diderot is under active development, and compatibility across versions is
not guaranteed: the MCP tools, the HTTP API, schemas, and settings change
together, and breaking changes are called out in the release notes
(Upgrading).
Security
Please report security issues privately, as described in
SECURITY.md: email [email protected]. Don't open a public
issue.
License
This project is dual-licensed under MIT and Apache 2.0 terms:
- MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)
- Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
Downstream projects and end users may choose either license individually, or both together, at their discretion. The motivation for this dual-licensing is the additional software patent assurance provided by Apache 2.0.
Bluesky Social PBC has committed to a software patent non-aggression pledge. For details see the original announcement.
Yorumlar (0)
Yorum birakmak icin giris yap.
Yorum birakSonuc bulunamadi