A Practitioner's Guide to Graphs - Tim Ainge, Good Collective

What was stored

cues

111

whisperx 111

chunks

26

from 111 cues

keyframes

25

kept of 31 captured

frames with text

25

418 lines read

chapters

0

from the source metadata

keyframe bytes

1.9 MB

word timings on 111 cues

Provenance

Frames, and what the machine read

• 

  0:10 #0 done6 line(s)

  shot 0·sharpness 319.8

  1. AlEngineer0.96
  2. World'sFair1.00
  3. A Practitioner's Guide to Graphs1.00
  4. How to make you Al applications smarter, cheaper, and more reliable0.99
  5. GoodCollective1.00
  6. CapabilityxStrategy=Delivery1.00
• 

  0:40 #1 done3 line(s)

  shot 1·sharpness 795.6

  1. If I have a hammer, is everything a graph?1.00
  2. Obsidian knowledge graph1.00
  3. GraphRA60.95
• 

  1:11 #2 done6 line(s)

  shot 2·sharpness 899.9

  1. If I have a hammer, is everything a graph?1.00
  2. graphs from1.00
  3. 1st principles1.00
  4. Obsidian knowledge graph1.00
  5. A0.52
  6. GraphRA60.94
• 

  1:21 #3 skipped

  shot 3·duplicate of #2

• 

  1:56 #4 done10 line(s)

  shot 4·sharpness 520.3

  1. The structure of things to come1.00
  2. The talk1.00
  3. PART1.00
  4. Personalised PageRank1.00
  5. COVERS1.00
  6. Graph algorithms0.99
  7. COVERS1.00
  8. Shortest path0.99
  9. COVERS1.00
  10. Subgraph matching1.00
• 

  2:01 #5 done4 line(s)

  shot 5·sharpness 317.8

  1. The structure of things to come0.99
  2. Principle1.00
  3. Code1.00
  4. Examples1.00
• 

  2:22 #6 done14 line(s)

  shot 6·sharpness 274.8

  1. What is a graph1.00
  2. Nodes, edges, weights, properties.1.00
  3. Milk0.97
  4. CONTAINS1.00
  5. Whisk together1.00
  6. HAS_STEP1.00
  7. Pancakes1.00
  8. CONTAINS1.00
  9. CONTAINS1.00
  10. Recipe1.00
  11. Egg1.00
  12. Ingredient1.00
  13. Flour1.00
  14. Step1.00
• 

  2:51 #7 done37 line(s)

  shot 7·sharpness 661.7

  1. Extract a basic graph0.99
  2. Principle: unstructured data isn't much more helpful than unstructured text.1.00
  3. from pydantic import BaseModel1.00
  4. 4 people serves0.99
  5. from pydantic_ai import Agent1.00
  6. pancake recipe0.99
  7. requires1.00
  8. class Triple(BaseModel):1.00
  9. needs1.00
  10. subject: str0.97
  11. predicate: str1.00
  12. plain fid@pbined with0.88
  13. object: str1.00
  14. Milk1.00
  15. contains1.00
  16. # no schema, no vocabulary, no entity resolution0.99
  17. breakfast1.00
  18. agent = Agent(0.98
  19. is a0.92
  20. "anthropic:claude-sonnet-4-6",1.00
  21. Pancakes1.00
  22. has ingredient1.00
  23. output_type=list[Triple],1.00
  24. eggs1.00
  25. system_prompt=(1.00
  26. needs1.00
  27. "Pull the key facts out as (subject, predicate,1.00
  28. "object) triples. Use whatever vocab fits, yolo."1.00
  29. ),0.94
  30. Flour1.00
  31. made withixed with0.99
  32. triples = agent.run_sync(recipe_text).output0.99
  33. cooked on0.97
  34. uses1.00
  35. the batter0.95
  36. Egg1.00
  37. a pan0.99
• 

  3:23 #8 skipped

  shot 8·duplicate of #7

• 

  4:13 #9 done24 line(s)

  shot 9·sharpness 700.6

  1. Defining the schema (the shape)1.00
  2. Principle: give the extractor a shape - a schema - to fill. A shape gets you structure.0.98
  3. from pydantic import BaseModel, Field0.98
  4. from pydantic_ai import Agent1.00
  5. garlic1.00
  6. class Ingredient(BaseModel):1.00
  7. name: str1.00
  8. quantity: float | None = None0.99
  9. unit: str |0.97
  10. None = None0.99
  11. # 'cup', 'g', 'tbsp'0.97
  12. parsley1.00
  13. parmesan1.00
  14. class Recipe(BaseModel):0.99
  15. title: str0.97
  16. ingredients: list[Ingredient]0.99
  17. Garlic Butter Pasta1.00
  18. CONTAINS1.00
  19. # the schema IS the contract - no free-form vocab0.99
  20. ...0.94
  21. agent = Agent("anthropic:claude-sonnet-4-6", output_type=Recipe)0.99
  22. recipe = agent.run_sync(recipe_text).output1.00
  23. spaghetti1.00
  24. butter1.00
• 

  4:30 #10 done40 line(s)

  shot 10·sharpness 1022.4

  1. A recipe is more than its ingredients0.98
  2. The shape grows with the domain - add steps , and the technique each one applies.0.98
  3. toss1.00
  4. class Ingredient(BaseModel):1.00
  5. name: str0.98
  6. TECHNIQUE1.00
  7. quantity: float | None = Field(None, description="...")0.97
  8. unit: str |0.95
  9. None = Field(None, description="... ")0.96
  10. Toss the pasta in tese's0.87
  11. parsley1.00
  12. class Step(BaseModel):1.00
  13. HAS_STEP1.00
  14. text: str1.00
  15. parmesan1.00
  16. CONTAINS1.00
  17. technique: str | None = Field(None, description="... ")0.97
  18. spaghetti1.00
  19. CONTAINS1.00
  20. CONTAINS1.00
  21. uses: list[str] = Field(default_factory=list, description=".0.99
  22. CONTAINS1.00
  23. garlic1.00
  24. USES1.00
  25. HAS_STEP1.00
  26. Garlic Butter Pasta1.00
  27. HAS_STEP0.99
  28. CONTAINS0.99
  29. class Recipe(BaseModel):1.00
  30. title: str0.99
  31. TECHNIQUE0.99
  32. Boil the spaghetti un...0.99
  33. Melt the butter, then...0.95
  34. steps: list[Step]0.99
  35. ingredients: list[Ingredient]0.99
  36. boil0.95
  37. butter0.98
  38. TECHNIQUE1.00
  39. saute1.00
  40. + Techniques - the verb each step applies (saute, boil, toss)0.98
• 

  4:44 #11 done42 line(s)

  shot 11·sharpness 1132.6

  1. Adding ontology in the prompt0.99
  2. Principle: a schema provides shape, the ontology describes what should go into it.0.99
  3. Oyster Sauce0.99
  4. Salt1.00
  5. # the ontology, expressed as rules in the system prompt0.99
  6. Oil1.00
  7. Minced Garlic1.00
  8. SYSTEM_PROMPT = (0.98
  9. Bean Sprouts0.98
  10. 1 tbs pinch0.97
  11. "Extract recipes into STRICT, schema-conformant form:\n"0.99
  12. 5 tbs0.99
  13. 1 tsp0.97
  14. Egg1.00
  15. "1. CANONICAL UNITS - convert to SI:\n"0.99
  16. 1 cup0.96
  17. mass → grams ('g'); volume → millilitres ('ml')\n"0.99
  18. Sesame Seed Oil1.00
  19. 2 tsp0.99
  20. 0.51.00
  21. (1 cup → 240 ml; 1 tbsp → 15 ml; 1 oz → 28 g)\n"0.97
  22. Mushrooms1.00
  23. countable items (eggs, cloves) keep a null unit.\n"1.00
  24. CANONICAL NAMES - clean, lowercase, singular:\n"0.99
  25. 0.25 lb0.97
  26. 'plain flour' / 'all-purpose' → 'flour'\n"0.98
  27. Soy Sauce1.00
  28. 0.5 cup1.00
  29. Beef Lo Mein1.00
  30. Noodles1.00
  31. strip brands & loose adjectives ('chopped', 'fresh'),\n"1.00
  32. 0.5 lb0.93
  33. keep distinguishing words ('smoked', 'self-raising').\n"1.00
  34. "Conform to the schema; don't invent ingredients."1.00
  35. Onion1.00
  36. 3 tbs1.00
  37. Sugar1.00
  38. Beef1.00
  39. Pepper1.00
  40. Water1.00
  41. Ginger1.00
  42. Starch1.00
• 

  6:02 #12 done30 line(s)

  shot 12·sharpness 900.5

  1. Simple entity matching1.00
  2. Principle: epic power-up = resolve your entities.0.99
  3. CONTAINS0.98
  4. soy sauce0.99
  5. Beef Lo Mein0.99
  6. # a synonym / index table maps surface forms → one canonical nod0.99
  7. INGREDIENTS = {0.97
  8. CONTAINS1.00
  9. CONTAINSCONTAINS0.99
  10. "garlic": ["garlic clove", "minced garlic", "2 cloves garlic"1.00
  11. CONTAINS1.00
  12. water1.00
  13. "shallot": ["challots", "eschalot", "golden shallot"],0.99
  14. Teriyaki Chicken CasS.CONTANS0.95
  15. concaunSpiced Fish Tac...0.81
  16. 10.65
  17. garlic1.00
  18. def normalise(name: str) → str:0.98
  19. onion1.00
  20. cleaned = clean(name)0.99
  21. # lowercase, str0.97
  22. CONT1.00
  23. CONTANS0.94
  24. EONTAINSCONTAINS0.96
  25. return SYN_TO_CANON.get(cleaned, cleaned)0.99
  26. # exact hit → ca0.94
  27. Egyptian Fatteh0.98
  28. Chicken Handi0.99
  29. CONTAINS1.00
  30. cumin1.00
• 

  6:22 #13 done20 line(s)

  shot 13·sharpness 647.5

  1. Better entity matching1.00
  2. Principle: Hybrid approaches often yield the best results. In this case semantic entity matching.1.00
  3. # hybrid match = semantic vector score + lexical score0.98
  4. def hybrid_lookup(query, candidates):0.99
  5. q, *cands = embed([query, *candidates])0.99
  6. # all-MiniLM-L6-v20.99
  7. coriander1.00
  8. capsicurhilli1.00
  9. scores = []0.99
  10. apoing onion0.96
  11. for cand, vec in zip(candidates, cands):0.99
  12. flour0.98
  13. vector = (cosine(q, vec) + 1) / 20.97
  14. # semantic kinship1.00
  15. lexical = token_sort_ratio(query, cand) / 100 # surface overlap0.99
  16. milk1.00
  17. sugar1.00
  18. scores.append(0.6 * vector + 0.4 * lexical)0.99
  19. return best_first(candidates, scores)1.00
  20. hybrid_lookup("garbanzo", existing_nodes) # → 'chickpea'0.97
• 

  6:43 #14 done15 line(s)

  shot 14·sharpness 412.0

  1. Graph queries1.00
  2. Principle: query by relationship, not just by node – and get a subgraph back.0.99
  3. // graph - walk the relationship0.95
  4. MATCH (r:Recipe)-[:CONTAINS]→(i:Ingredient)0.99
  5. WHERE i.name = 'garlic'0.97
  6. RETURN r,i0.99
  7. -- relational - join through the link table0.99
  8. garlic?1.00
  9. SELECT r.title1.00
  10. FROM recipes r0.99
  11. JOIN recipe_ingredients ri ON ri.recipe_id = r.id1.00
  12. JOIN ingredients0.98
  13. i ON i.id = ri.ingredient_id1.00
  14. WHERE i.name = 'garlic';1.00
  15. garlic?– the question, as a seed node0.98
• 

  7:22 #15 done34 line(s)

  shot 15·sharpness 795.2

  1. Graph queries1.00
  2. Principle: query by relationship, not just by node – and get a subgraph back.0.99
  3. cannellini beans0.99
  4. CONTAINS1.00
  5. // graph - walk the relationship0.97
  6. MATCH (r:Recipe)-[:CONTAINS]→(i:Ingredient)0.99
  7. Croatian Bean Stew0.99
  8. WHERE i.name = 'garlic'0.98
  9. CONTAINS0.98
  10. CONTAINS1.00
  11. RETURN r, i0.96
  12. beef0.96
  13. CONTAINS0.99
  14. Egyptian Fatteh1.00
  15. CONTAINS1.00
  16. -- relational - join through the link table0.99
  17. CONTAINS0.95
  18. garlic?1.00
  19. CONTAINS1.00
  20. SELECT r.title1.00
  21. FROM recipes r0.99
  22. Beef Lo Mein1.00
  23. CONTAINS1.00
  24. Chicken Handi1.00
  25. CONTAINS1.00
  26. onion1.00
  27. JOIN recipe_ingredients ri ON ri.recipe_id = r.id0.99
  28. JOIN ingredients0.99
  29. i ON i.id = ri.ingredient_id0.98
  30. WHERE i.name = 'garlic';1.00
  31. Cajun Spiced Fish Tac...1.00
  32. CONTAINS0.97
  33. white fish1.00
  34. → + their other ingredients - the subgraph you'd hand a model0.97
• 

  7:41 #16 done4 line(s)

  shot 16·sharpness 267.4

  1. Personalised PageRank0.99
  2. Plainly: pick one node, wander the edges, keep resetting to where you started. Where you land most often is0.99
  3. what's relevant.0.98
  4. ChickenHandi1.00
• 

  8:21 #17 done4 line(s)

  shot 17·sharpness 288.1

  1. Personalised PageRank1.00
  2. Plainly: pick one node, wander the edges, keep resetting to where you started. Where you land most often is0.99
  3. what's relevant.0.99
  4. Egyptian Fatteh0.96
• 

  8:41 #18 skipped

  shot 18·duplicate of #17

• 

  9:01 #19 done5 line(s)

  shot 19·sharpness 306.2

  1. Personalised PageRank1.00
  2. Principle: The most value comes from being able to highlight related nodes when they aren't obvious.1.00
  3. ginger0.99
  4. garlic1.00
  5. Add the onions and0.98
• 

  9:13 #20 done2 line(s)

  shot 20·sharpness 262.9

  1. Personalised PageRank0.99
  2. Principle: The most value comes from being able to highlight related nodes when they aren't obvious.1.00
• 

  9:32 #21 done26 line(s)

  shot 21·sharpness 1018.9

  1. PPR in the wild – landmark law0.97
  2. Real-world: seed a walk at one routine case; out of 27,885, the landmark it stands on surfaces.1.00
  3. Johnsonv.Zerbst1.00
  4. Malloy v.Hogan0.97
  5. cites1.00
  6. cites1.00
  7. Buchananv.Kentucky0.98
  8. Powell v Texas0.89
  9. Powell v.Alabama0.98
  10. cites1.00
  11. Estellev.Smith1.00
  12. cites1.00
  13. cites1.00
  14. cites1.00
  15. cites1.00
  16. cites1.00
  17. Kansas v. Cheever0.96
  18. cites1.00
  19. Massiahv.UnitedStates1.00
  20. Fitzpatrickv.UnitedStates1.00
  21. UnitedStatesv.Wade1.00
  22. cites1.00
  23. cites1.00
  24. Kansasv.Marsh0.99
  25. Brownv.UnitedStates1.00
  26. one hop further – into the precedent cluster0.97
• 

  10:07 #22 done22 line(s)

  shot 22·sharpness 501.3

  1. Shortest path1.00
  2. Principle: "how does A relate to B, and through what?"0.98
  3. # the path IS the explanation - one NetworkX call1.00
  4. path = nx.shortest_path(0.99
  5. code_graph,1.00
  6. # the eShop "calls" graph0.99
  7. source="CheckoutModel.OnPost",1.00
  8. # the symptom1.00
  9. target="Basket..ctor",1.00
  10. # the constructor we touched0.99
  11. BasketViewModelServic1.00
  12. → CheckoutModel.OnPost1.00
  13. CheckoutModel.SetBask..1.00
  14. #1.00
  15. CheckoutModel.SetBasketModelAsync1.00
  16. CheckoutModel.OnPost1.00
  17. BasketViewModelService.GetOrCreateBasketForUser1.00
  18. #1.00
  19. BasketViewModelService.CreateBasketForUser1.00
  20. #1.00
  21. Basket..ctor1.00
  22. the path is the explanation - a 4-hop calls chain from checkout down to the constructor0.99
• 

  10:32 #23 done25 line(s)

  shot 23·sharpness 566.1

  1. Shortest path1.00
  2. Principle: "how does A relate to B, and through what?"0.99
  3. # the path IS the explanation - one NetworkX call0.99
  4. path = nx.shortest_path(1.00
  5. code_graph,1.00
  6. # the eShop "calls" graph0.98
  7. source="CheckoutModel.OnPost",1.00
  8. # the symptom1.00
  9. target="Basket..ctor",1.00
  10. # the constructor we touched0.99
  11. BasketViewModelServic1.00
  12. → CheckoutModel.OnPost0.99
  13. CheckoutModel.SetBask..1.00
  14. #1.00
  15. #1.00
  16. → BasketViewModelService.GetOrCreateBasketForUser0.99
  17. CheckoutModel.SetBasketModelAsync1.00
  18. .ctor1.00
  19. BasketViewModelServic...0.96
  20. CheckoutModel.OnPost1.00
  21. #1.00
  22. BasketViewModelService.CreateBasketForUser0.99
  23. #1.00
  24. Basket..ctor0.98
  25. the path is the explanation - a 4-hop calls chain from checkout down to the constructor0.99

Transcript

111 cues· 2,277 words· 12,769 chars