Amr Gharbeia
717d63d84a
Some checks failed
Deploy (Gitea) / deploy (push) Failing after 3s
- Add mark-vector-stale(id, content): sets :vector to :pending, queues for re-embed - Register cron job (embed-all-pending every 10m on :reflex tier via orchestrator) - Add defskill :passepartout-system-model-embedding (priority 70) - Remove embedding from topological-sort exclusion list in core-skills - Export mark-vector-stale in core-defpackage - Contract: items 4 (mark-vector-stale) and 5 (cron registration) - Test: test-mark-vector-stale (5 checks) - ROADMAP: mark Context Manager, Async Embedding Gateway, TUI Experience as DONE - All v0.3.0 items now complete. Total: 5 suites, 85 checks, 0 failures
248 lines
10 KiB
Org Mode
248 lines
10 KiB
Org Mode
#+TITLE: SKILL: Embedding Gateway (org-skill-embedding-gateway.org)
|
|
#+AUTHOR: Agent
|
|
#+FILETAGS: :skill:system:embedding:
|
|
#+PROPERTY: header-args:lisp :tangle ../lisp/system-model-embedding.lisp
|
|
|
|
* Architectural Intent
|
|
|
|
~system-model-embedding~ converts text into vector representations for semantic search and memory retrieval. It provides three backends:
|
|
|
|
- ~:local~ — any OpenAI-compatible ~/api/embeddings~ endpoint (Ollama, vLLM, etc.)
|
|
- ~:openai~ — the OpenAI ~/v1/embeddings~ API with an API key
|
|
- ~:hashing~ — a zero-dependency fallback that produces deterministic vectors from SHA-256 hashes. No server, no config, works offline.
|
|
|
|
The embedding queue (~embed-queue-object~ / ~embed-all-pending~) decouples document indexing from the main loop. On each heartbeat tick, ~embed-all-pending~ drains the queue and embeds all accumulated objects. This prevents indexing traffic from blocking conversational responses.
|
|
|
|
The default provider is ~:hashing~ — useful for bootstrapping with zero configuration and for deployments where embedding quality isn't critical. Switch to ~:local~ or ~:openai~ when you have an embedding server available.
|
|
|
|
This replaces the old ~system-embedding-gateway~ with the same logic but renamed to ~system-model-embedding~ to live alongside the other ~system-model-*~ skills.
|
|
|
|
* Implementation
|
|
|
|
** State
|
|
#+begin_src lisp
|
|
(in-package :passepartout)
|
|
|
|
(defvar *embedding-provider* :hashing
|
|
"Active embedding provider: :hashing, :local, :openai.")
|
|
|
|
(defvar *embedding-queue* nil
|
|
"Queue of text objects awaiting embedding.")
|
|
|
|
(defvar *embedding-batch-size* 10
|
|
"Maximum texts per embedding API call.")
|
|
#+end_src
|
|
|
|
** Local backend (OpenAI-compatible)
|
|
#+begin_src lisp
|
|
(defun embedding-backend-local (text)
|
|
"Generate embeddings via a local OpenAI-compatible endpoint."
|
|
(let* ((url (or (uiop:getenv "LOCAL_BASE_URL") (format nil "http://~a" (or (uiop:getenv "OLLAMA_HOST") "localhost:11434"))))
|
|
(model (or (uiop:getenv "EMBEDDING_MODEL") "nomic-embed-text"))
|
|
(body (cl-json:encode-json-to-string
|
|
`((model . ,model) (input . ,text)))))
|
|
(handler-case
|
|
(let* ((response (dex:post (format nil "~a/api/embeddings" url)
|
|
:headers '(("Content-Type" . "application/json"))
|
|
:content body :connect-timeout 5 :read-timeout 30))
|
|
(json (cl-json:decode-json-from-string response))
|
|
(data (car (cdr (assoc :data json)))))
|
|
(or (cdr (assoc :embedding data))
|
|
(list :error "No embedding in response")))
|
|
(error (c)
|
|
(list :error (format nil "Embedding failed: ~a" c))))))
|
|
#+end_src
|
|
|
|
** OpenAI backend
|
|
#+begin_src lisp
|
|
(defun embedding-backend-openai (text)
|
|
"Generate embeddings via OpenAI compatible /v1/embeddings endpoint."
|
|
(let* ((api-key (uiop:getenv "OPENAI_API_KEY"))
|
|
(base-url (or (uiop:getenv "EMBEDDING_BASE_URL") "https://api.openai.com/v1"))
|
|
(model (or (uiop:getenv "EMBEDDING_MODEL") "text-embedding-3-small"))
|
|
(body (cl-json:encode-json-to-string
|
|
`((model . ,model) (input . ,text)))))
|
|
(handler-case
|
|
(let* ((response (dex:post (format nil "~a/embeddings" base-url)
|
|
:headers `(("Content-Type" . "application/json")
|
|
("Authorization" . ,(format nil "Bearer ~a" api-key)))
|
|
:content body :connect-timeout 5 :read-timeout 30))
|
|
(json (cl-json:decode-json-from-string response))
|
|
(data (car (cdr (assoc :data json)))))
|
|
(or (cdr (assoc :embedding data))
|
|
(list :error "No embedding in response")))
|
|
(error (c)
|
|
(list :error (format nil "OpenAI Embedding failed: ~a" c))))))
|
|
#+end_src
|
|
|
|
** Hashing fallback
|
|
#+begin_src lisp
|
|
(defun embedding-backend-hashing (text)
|
|
"Fallback: produces a deterministic vector from the text hash."
|
|
(let* ((digest (ironclad:digest-sequence :sha256 (babel:string-to-octets text)))
|
|
(vec (make-array 8 :element-type 'single-float :initial-element 0.0)))
|
|
(dotimes (i (min (length digest) 8))
|
|
(setf (aref vec i) (float (/ (aref digest i) 255.0) 0.0)))
|
|
vec))
|
|
#+end_src
|
|
|
|
** Object embedding and queuing
|
|
#+begin_src lisp
|
|
(defvar *embedding-backend* nil
|
|
"Explicit backend override (nil = use *embedding-provider*).")
|
|
|
|
(defun embeddings-compute (text)
|
|
"Compute an embedding vector for text using the active backend."
|
|
(embed-object text))
|
|
|
|
(defun embed-object (text)
|
|
"Embed a single text string using the active backend."
|
|
(let* ((selected (or *embedding-backend* *embedding-provider* :hashing))
|
|
(backend (case selected
|
|
(:local #'embedding-backend-local)
|
|
(:openai #'embedding-backend-openai)
|
|
(t #'embedding-backend-hashing))))
|
|
(if backend
|
|
(progn
|
|
(log-message "EMBEDDING: Provider ~a, backend=~a" selected backend)
|
|
(funcall backend text))
|
|
(progn
|
|
(log-message "EMBEDDING: No backend for provider ~a, using hashing" selected)
|
|
(embedding-backend-hashing text)))))
|
|
|
|
(defun embed-queue-object (object)
|
|
"Queue a text object for async embedding."
|
|
(push object *embedding-queue*)
|
|
(log-message "EMBEDDING: Queued object"))
|
|
|
|
(defun embed-all-pending ()
|
|
"Drain the embedding queue, store vectors in the store-keyed objects."
|
|
(let ((batch (nreverse *embedding-queue*)))
|
|
(setf *embedding-queue* nil)
|
|
(dolist (item batch)
|
|
(handler-case
|
|
(let ((id (getf item :id))
|
|
(text (getf item :text)))
|
|
(when (and id text)
|
|
(let ((vec (embeddings-compute text))
|
|
(obj (gethash id *memory-store*)))
|
|
(when (and obj vec (not (listp vec)))
|
|
(setf (memory-object-vector obj) vec))
|
|
(log-message "EMBEDDING: Computed vector for ~a (~d dims)" id (length vec)))))
|
|
(error (c)
|
|
(log-message "EMBEDDING: Failed to embed object: ~a" c))))))
|
|
|
|
;; Apply env var override at load time
|
|
(let ((provider-env (uiop:getenv "EMBEDDING_PROVIDER")))
|
|
(when provider-env
|
|
(let ((kw (intern (string-upcase provider-env) :keyword)))
|
|
(setf *embedding-provider* kw)
|
|
(log-message "EMBEDDING: Set provider to ~a from EMBEDDING_PROVIDER env" kw))))
|
|
|
|
(log-message "EMBEDDING: Gateway loaded with provider ~a" *embedding-provider*)
|
|
#+end_src
|
|
|
|
** Stale vector marking
|
|
#+begin_src lisp
|
|
(defun mark-vector-stale (id &optional content)
|
|
"Mark a memory object's vector as :pending and queue it for re-embedding.
|
|
When content is not supplied, reads from the object in *memory-store*."
|
|
(let* ((obj (gethash id *memory-store*))
|
|
(text (or content (and obj (memory-object-content obj)))))
|
|
(when obj
|
|
(setf (memory-object-vector obj) :pending))
|
|
(when text
|
|
(push (list :id id :text text) *embedding-queue*)
|
|
(log-message "EMBEDDING: Marked ~a vector stale, queued for re-embed" id))
|
|
(or obj text)))
|
|
#+end_src
|
|
|
|
** Skill Registration and Cron
|
|
#+begin_src lisp
|
|
(defskill :passepartout-system-model-embedding
|
|
:priority 70
|
|
:trigger (lambda (ctx) (declare (ignore ctx)) nil))
|
|
|
|
;; Register periodic batch embedding via cron (when orchestrator available)
|
|
(when (fboundp 'orchestrator-register-cron)
|
|
(handler-case
|
|
(orchestrator-register-cron :embed-batch
|
|
"<2026-05-05 Tue +10m>"
|
|
'embed-all-pending
|
|
:reflex)
|
|
(error (c)
|
|
(log-message "EMBEDDING: Cron registration failed: ~a" c))))
|
|
#+end_src
|
|
|
|
* Contract
|
|
|
|
1. (embeddings-compute text): produces a vector (single-float array) for
|
|
any text string using the active backend (~*embedding-backend*~ or
|
|
~*embedding-provider*~).
|
|
2. (embedding-backend-hashing text): zero-dependency fallback. Returns
|
|
an 8-element single-float vector deterministically from SHA-256.
|
|
3. (embed-all-pending): drains ~*embedding-queue*~, computes vectors for
|
|
all queued objects, and stores them in ~*memory-store*~ entries.
|
|
4. (mark-vector-stale id &optional content): sets ~:vector~ to ~:pending~
|
|
and pushes object to ~*embedding-queue*~ for background re-embedding.
|
|
5. Cron: ~embed-all-pending~ is registered with the orchestrator to run
|
|
on ~:reflex~ tier every 10 minutes for background batch processing.
|
|
|
|
* Test Suite
|
|
#+begin_src lisp
|
|
(eval-when (:compile-toplevel :load-toplevel :execute)
|
|
(ql:quickload :fiveam :silent t))
|
|
|
|
(defpackage :passepartout-embedding-tests
|
|
(:use :cl :passepartout)
|
|
(:export #:embedding-suite))
|
|
|
|
(in-package :passepartout-embedding-tests)
|
|
|
|
(fiveam:def-suite embedding-suite :description "Embedding gateway verification")
|
|
(fiveam:in-suite embedding-suite)
|
|
|
|
(fiveam:test test-embedding-backend-hashing
|
|
"Contract 2: hashing backend produces 8-element float vector."
|
|
(let ((vec (embedding-backend-hashing "hello world")))
|
|
(fiveam:is (arrayp vec))
|
|
(fiveam:is (= 8 (length vec)))
|
|
(fiveam:is (every #'numberp (coerce vec 'list)))))
|
|
|
|
(fiveam:test test-embedding-backend-hashing-deterministic
|
|
"Contract 2: same input produces same vector."
|
|
(let ((v1 (embedding-backend-hashing "test"))
|
|
(v2 (embedding-backend-hashing "test")))
|
|
(fiveam:is (equalp v1 v2))))
|
|
|
|
(fiveam:test test-embeddings-compute
|
|
"Contract 1: embeddings-compute returns a float vector."
|
|
(let ((vec (embeddings-compute "some text")))
|
|
(fiveam:is (arrayp vec))
|
|
(fiveam:is (> (length vec) 0))))
|
|
|
|
(fiveam:test test-embed-queue-and-drain
|
|
"Contract 3: embed-all-pending drains queue and stores vectors."
|
|
(let ((*embedding-queue* nil))
|
|
(embed-queue-object '(:id "test-obj" :text "sample text"))
|
|
(fiveam:is (= 1 (length *embedding-queue*)))
|
|
(embed-all-pending)
|
|
(fiveam:is (null *embedding-queue*))))
|
|
|
|
(fiveam:test test-mark-vector-stale
|
|
"Contract 4: mark-vector-stale sets vector to :pending and queues for re-embed."
|
|
(let ((*embedding-queue* nil))
|
|
;; Create an object in memory with a vector
|
|
(let ((obj (make-memory-object :id "stale-test" :content "stale content"
|
|
:vector #(1.0 2.0 3.0))))
|
|
(setf (gethash "stale-test" *memory-store*) obj)
|
|
(mark-vector-stale "stale-test")
|
|
(fiveam:is (eq :pending (memory-object-vector obj)))
|
|
(fiveam:is (= 1 (length *embedding-queue*)))
|
|
(let ((item (first *embedding-queue*)))
|
|
(fiveam:is (string= "stale-test" (getf item :id)))
|
|
(fiveam:is (string= "stale content" (getf item :text))))
|
|
;; Clean up
|
|
(remhash "stale-test" *memory-store*))))
|
|
#+end_src
|