A Clojure library inspired by DSPy, providing a declarative approach to building and optimizing language model pipelines.
DSCloj leverages litellm-clj to provide a unified interface for working with various LLM providers while bringing DSPy's powerful programming model to the Clojure ecosystem.
DSCloj brings the power of declarative LLM programming to Clojure. Inspired by Stanford's DSPy framework.
Add DSCloj to your deps.edn:
{:deps {io.unravel/dscloj {:mvn/version "0.1.0"}}}
DSCloj works by defining modules - declarative specifications of LLM tasks with typed inputs and outputs.
(require '[dscloj.core :as dscloj])
;; 1. Define a module with Malli specs
(def qa-module
{:inputs [{:name :question
:spec :string
:description "The question to answer"}]
:outputs [{:name :answer
:spec :string
:description "The answer to the question"}]
:instructions "Provide concise and accurate answers."})
;; 2. Use the module with predict
(def result (dscloj/predict qa-module
{:question "What is the capital of France?"}
{:model "gpt-4"
:api-key (System/getenv "OPENAI_API_KEY")}))
;; 3. Access the structured output
(:answer result)
;; => "Paris"
Modules are maps with:
:inputs - Vector of input field definitions:outputs - Vector of output field definitions:instructions - Optional string describing the task instructions, rules, and examplesFields are maps with:
:name - Keyword identifier:spec - Malli spec (e.g., :string, :int, :double, :boolean, or more complex specs like [:int {:min 0 :max 100}]):description - Human-readable descriptionThe predict function:
DSCloj uses Malli specs for defining field types with automatic validation:
;; Simple specs
(def qa-module
{:inputs [{:name :question
:spec :string
:description "The question to answer"}]
:outputs [{:name :answer
:spec :string
:description "The answer"}]
:instructions "Provide concise and accurate answers."})
;; Complex specs with constraints
(def constrained-module
{:inputs [{:name :age
:spec [:int {:min 0 :max 150}]}
{:name :email
:spec [:string {:pattern #"^[^@]+@[^@]+$"}]}]
:outputs [{:name :message
:spec :string}]})
;; Reusable specs
(def QuestionSpec [:string {:min 1 :description "The question"}])
(def AnswerSpec [:string {:min 1 :description "The answer"}])
(def module-with-reusable-specs
{:inputs [{:name :question :spec QuestionSpec}]
:outputs [{:name :answer :spec AnswerSpec}]})
;; Invalid inputs/outputs are automatically validated
(dscloj/predict qa-module
{:question 123} ; Throws validation error - should be string
{:model "gpt-4"})
;; Disable validation if needed
(dscloj/predict qa-module
{:question "..."}
{:model "gpt-4"
:validate? false}) ; Skip validation
Benefits of Malli Specs:
DSCloj supports streaming structured output with progressive parsing and validation:
(require '[dscloj.core :as dscloj]
'[clojure.core.async :refer [go-loop <!]])
;; Define a module with structured outputs
(def whales-module
{:inputs [{:name :query :spec :string}]
:outputs [{:name :species_1_name :spec :string}
{:name :species_1_length :spec :double}
{:name :species_1_weight :spec :double}
;; ... more fields
]
:instructions "Generate whale species information."})
;; Stream predictions
(let [stream-ch (dscloj/predict-stream
whales-module
{:query "Tell me about 3 whale species."}
{:model "gpt-4"
:api-key (System/getenv "OPENAI_API_KEY")
:debounce-ms 100})]
;; Consume the stream progressively
(go-loop []
(when-let [parsed (<! stream-ch)]
(println "Received update:" parsed)
(recur))))
Streaming Features:
See examples/streaming_whales.clj for a complete example inspired by Pydantic AI's streaming example.
See examples/basic_usage.clj for:
See examples/streaming_whales.clj for:
DSCloj uses litellm-clj and supports:
This project is licensed under the MIT License - see the LICENSE file for details.
Can you improve this documentation?Edit on GitHub
cljdoc builds & hosts documentation for Clojure/Script libraries
| Ctrl+k | Jump to recent docs |
| ← | Move to previous article |
| → | Move to next article |
| Ctrl+/ | Jump to the search field |