dspy-ai-apps

Name: dspy-ai-apps
Author: One-Man-Company

77stars16forks

How to install dspy-ai-apps?

one-man-company/skills-contextmanager/dspy-ai-apps

$npx -y skills add One-Man-Company/Skills-ContextManager --skill dspy-ai-apps

Installs into the current project.

Summary

Comprehensive guide for building AI applications using the DSPy framework with Google's Gemini API. This skill enforces strict rate limit management for the 'gemini-3-flash-preview' model (20 RPD) and provides best practices for "Intent-Oriented Programming", decomposition, agent

Files · 1

View on GitHub

SKILL.md

1	# DSPy AI Applications with Gemini
2
3	This skill guides you through building robust, self-optimizing AI applications using DSPy (Declarative Self-improving Python) and the Gemini API. It emphasizes Intent-Oriented Programming: defining what you want (Signatures) rather than how to prompt it.
4
5	> [!WARNING]
6	> STRICT RATE LIMITS: The `gemini-3-flash-preview` model has extremely strict rate limits:
7	> - 5 Requests Per Minute (RPM)
8	> - 20 Requests Per Day (RPD)
9	> - 250,000 Tokens Per Minute (TPM)
10	>
11	> You MUST enable caching and use "Dry Runs" to avoid exhaustion.
12
13	## 1. Core Philosophy: Intent-Oriented Programming
14
15	DSPy shifts focus from "hand-crafting prompts" to "programming architectures".
16
17	1. Decomposition: Break complex tasks (Translation, RAG, Agents) into small, optimizable specific programs. Do not build monolithic "God Prompts".
18	2. Signatures: Declare input/output specs (including types like `list[dict]`, `float`, `Pydantic`).
19	3. Optimization: Use data (synthetic or real) to compiling programs into effective prompts automatically.
20	4. Modularity: Swap backends (e.g., Gemini -> Local) or optimization strategies without rewriting code.
21
22	## 2. Quick Start (Safe Mode)
23
24	To start a new project without hitting rate limits immediately:
25
26	1. Install Dependencies:
27	```bash
28	pip install dspy-ai google-generativeai
29	```
30
31	2. Use the Safe Boilerplate:
32	Always start with the provided boilerplate which includes caching and rate limit handling.
33	See [assets/boilerplate.py](assets/boilerplate.py).
34
35	3. Configure Environment:
36	Ensure `GOOGLE_API_KEY` is set in your environment variables.
37
38	## 3. Workflow
39
40	### Step 1: Define Typed Signatures
41	Define explicit intents. Be specific with types.
42	See [references/cheat_sheet.md](references/cheat_sheet.md) for examples (Invoice Parsing, Entity Extraction).
43
44	### Step 2: Build Modular Programs
45	Connect modules like `dspy.Predict`, `dspy.ChainOfThought`, or `dspy.ReAct`.
46	For Agents, expose these programs as tools.
47
48	### Step 3: Run Once & Cache
49	Run your module on a single example.
50	- ALWAYS use `dspy.configure(experimental=True)` or standard settings to enable file-based caching.
51	- Verify the cache file was created before proceeding.
52
53	### Step 4: Synthetic Data & Optimization (Advanced)
54	CRITICAL WARNING: Optimization loops (BootstrapFewShot, MIPRO) consume massive RPD.
55	- Strategy: Use a "Teacher" model (stronger/different quota) to generate synthetic data if possible.
56	- Micro-Optimization: If you must optimize on Gemini Flash, use a tiny trainset (2-3 examples) and `BootstrapFewShot` with `max_bootstrapped_demos=1`.
57
58	## 4. Advanced Patterns
59
60	### A. Agents & MCP
61	DSPy programs can be exposed as MCP (Model Context Protocol) tools.
62	1. Define a `dspy.Signature` for the tool.
63	2. Wrap it in a `dspy.Predict`.
64	3. Serve it via an MCP server (e.g., using `mcp2py`).
65
66	### B. Fine-Tuning Flow
67	1. Define Signature.
68	2. Optimize a Teacher program (strong model) to get high-quality traces.
69	3. Generate synthetic data.
70	4. Fine-tune a smaller Student model (e.g., Gemma 2B) using `BootstrapFinetune`.
71
72	### C. RAG (Retrieval Augmented Generation)
73	Combine `dspy.Retrieve` (e.g., ColBERTv2, VectorDB) with `dspy.ChainOfThought`.
74	Optimize the entire pipeline to improve retrieval queries and answer generation simultaneously.
75
76	## 5. Troubleshooting
77
78	- 429 Errors: You hit the rate limit. Stop immediately.
79	- Empty Responses: Check API key and safety settings.
80	- "Context too long": Use `dspy.Retrieve` or decompose the task.
81
82	## 6. Artifacts & Resources
83
84	- `assets/boilerplate.py`: MANDATORY starting point (includes Pydantic/Typed examples).
85	- `references/cheat_sheet.md`: Signatures for Invoice Parser, RAG, Agents, Entities.

People who liked this also viewed

6 picks

Type

dspy-ai-apps

77stars16forks

View on GitHub

How to install dspy-ai-apps?

one-man-company/skills-contextmanager/dspy-ai-apps

$npx -y skills add One-Man-Company/Skills-ContextManager --skill dspy-ai-apps

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# DSPy AI Applications with Gemini
2
3	This skill guides you through building robust, self-optimizing AI applications using DSPy (Declarative Self-improving Python) and the Gemini API. It emphasizes Intent-Oriented Programming: defining what you want (Signatures) rather than how to prompt it.
4
5	> [!WARNING]
6	> STRICT RATE LIMITS: The `gemini-3-flash-preview` model has extremely strict rate limits:
7	> - 5 Requests Per Minute (RPM)
8	> - 20 Requests Per Day (RPD)
9	> - 250,000 Tokens Per Minute (TPM)
10	>
11	> You MUST enable caching and use "Dry Runs" to avoid exhaustion.
12
13	## 1. Core Philosophy: Intent-Oriented Programming
14
15	DSPy shifts focus from "hand-crafting prompts" to "programming architectures".
16
17	1. Decomposition: Break complex tasks (Translation, RAG, Agents) into small, optimizable specific programs. Do not build monolithic "God Prompts".
18	2. Signatures: Declare input/output specs (including types like `list[dict]`, `float`, `Pydantic`).
19	3. Optimization: Use data (synthetic or real) to compiling programs into effective prompts automatically.
20	4. Modularity: Swap backends (e.g., Gemini -> Local) or optimization strategies without rewriting code.
21
22	## 2. Quick Start (Safe Mode)
23
24	To start a new project without hitting rate limits immediately:
25
26	1. Install Dependencies:
27	```bash
28	pip install dspy-ai google-generativeai
29	```
30
31	2. Use the Safe Boilerplate:
32	Always start with the provided boilerplate which includes caching and rate limit handling.
33	See [assets/boilerplate.py](assets/boilerplate.py).
34
35	3. Configure Environment:
36	Ensure `GOOGLE_API_KEY` is set in your environment variables.
37
38	## 3. Workflow
39
40	### Step 1: Define Typed Signatures
41	Define explicit intents. Be specific with types.
42	See [references/cheat_sheet.md](references/cheat_sheet.md) for examples (Invoice Parsing, Entity Extraction).
43
44	### Step 2: Build Modular Programs
45	Connect modules like `dspy.Predict`, `dspy.ChainOfThought`, or `dspy.ReAct`.
46	For Agents, expose these programs as tools.
47
48	### Step 3: Run Once & Cache
49	Run your module on a single example.
50	- ALWAYS use `dspy.configure(experimental=True)` or standard settings to enable file-based caching.
51	- Verify the cache file was created before proceeding.
52
53	### Step 4: Synthetic Data & Optimization (Advanced)
54	CRITICAL WARNING: Optimization loops (BootstrapFewShot, MIPRO) consume massive RPD.
55	- Strategy: Use a "Teacher" model (stronger/different quota) to generate synthetic data if possible.
56	- Micro-Optimization: If you must optimize on Gemini Flash, use a tiny trainset (2-3 examples) and `BootstrapFewShot` with `max_bootstrapped_demos=1`.
57
58	## 4. Advanced Patterns
59
60	### A. Agents & MCP
61	DSPy programs can be exposed as MCP (Model Context Protocol) tools.
62	1. Define a `dspy.Signature` for the tool.
63	2. Wrap it in a `dspy.Predict`.
64	3. Serve it via an MCP server (e.g., using `mcp2py`).
65
66	### B. Fine-Tuning Flow
67	1. Define Signature.
68	2. Optimize a Teacher program (strong model) to get high-quality traces.
69	3. Generate synthetic data.
70	4. Fine-tune a smaller Student model (e.g., Gemma 2B) using `BootstrapFinetune`.
71
72	### C. RAG (Retrieval Augmented Generation)
73	Combine `dspy.Retrieve` (e.g., ColBERTv2, VectorDB) with `dspy.ChainOfThought`.
74	Optimize the entire pipeline to improve retrieval queries and answer generation simultaneously.
75
76	## 5. Troubleshooting
77
78	- 429 Errors: You hit the rate limit. Stop immediately.
79	- Empty Responses: Check API key and safety settings.
80	- "Context too long": Use `dspy.Retrieve` or decompose the task.
81
82	## 6. Artifacts & Resources
83
84	- `assets/boilerplate.py`: MANDATORY starting point (includes Pydantic/Typed examples).
85	- `references/cheat_sheet.md`: Signatures for Invoice Parser, RAG, Agents, Entities.

People who liked this also viewed

6 picks

Type