[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M71FA78Y5WJV6R6HGFSVT6XP":3,"skill-same:01M71FA78Y5WJV6R6HGFSVT6XP":85},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":39,"linkOnly":40,"markdownContent":41,"markdownBody":42,"frontMatter":43,"refName":53,"commitSha":54,"skillDirectory":55,"forksCount":56,"repositoryHtmlUrl":57,"repositoryHomepageUrl":6,"ownerHtmlUrl":58,"ownerBlogUrl":6,"ownerDescription":59,"updatedAt":60,"install":61,"resources":74},"01M71FA78Y5WJV6R6HGFSVT6XP","qdrant-cost-reduction-advisor",null,"Diagnoses when and why to recommend specific cost optimization techniques for Qdrant customers. Use when a customer or SA says \"too expensive\", \"need to reduce cost\", \"pricing too high\", \"how to lower price\", \"sticker shock\", or asks about scaling costs. Also trigger when a deal is blocked by pricing, a customer compares competitor pricing, a startup says the quote is too high, or someone needs a formula to predict costs at scale. This skill answers \"when should I recommend scalar vs binary quantization?\" not \"how to configure quantization.\" Do NOT use for general Qdrant setup, feature requests, or implementation how-tos.","qdrant-labs\u002Fdevrel-projects","qdrant-labs","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F269096346?v=4","01M71F98K9TZ3KQTNZBM116ZYG","Qdrant Labs","01M71F98M75QPSWAB0F1BZJZCN",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fqdrant-labs\u002Fdevrel-projects\u002Fblob\u002Fmain\u002Fqdrant-os\u002Fskills\u002Fqdrant-cost-reduction-advisor\u002FSKILL.md",[],[19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38],"qdrant cost reduction advisor","cost","Qdrant","advisor","feature","do","recommend","tos","at","why","how","customer","answers","configure","reduce","when","how to","formula","sticker","scaling","2026-11-10 10:12:30",false,"---\nname: qdrant-cost-reduction-advisor\ndescription: Diagnoses when and why to recommend specific cost\n  optimization techniques for Qdrant customers. Use when a customer\n  or SA says \"too expensive\", \"need to reduce cost\", \"pricing too\n  high\", \"how to lower price\", \"sticker shock\", or asks about scaling\n  costs. Also trigger when a deal is blocked by pricing, a customer\n  compares competitor pricing, a startup says the quote is too high,\n  or someone needs a formula to predict costs at scale. This skill\n  answers \"when should I recommend scalar vs binary quantization?\" not\n  \"how to configure quantization.\" Do NOT use for general Qdrant setup,\n  feature requests, or implementation how-tos.\nlicense: Apache-2.0\nmetadata:\n  author: thierry\n  version: 1.0.0\n  category: sales\n  tags: [cost, pricing, quantization, deals, scaling, ram]\n---\n\n# When and why to recommend cost reduction techniques\n\nThis skill converts SA knowledge into a navigation system over Qdrant\ndocumentation. It answers \"when?\" and \"why?\" for cost reduction, not \"how?\"\nEvery solution links to docs for implementation details.\n\n## Symptom routing\n\n| Customer signal | Go to | Load reference |\n|----------------|-------|----------------|\n| \"Too expensive\", sticker shock on proposal | Symptom 1 below | `references\u002Fcost-formula.md` |\n| \"How much will it cost at X scale?\" | Symptom 2 below | `references\u002Fcost-formula.md` |\n| \"Competitor is cheaper\" | Symptom 3 below | `references\u002Fcompetitive-pricing.md` |\n| Startup, budget-constrained | Symptom 4 below | `references\u002Fcost-formula.md` |\n\n## Do NOT recommend\n\n- Do NOT lead with \"Qdrant is cheaper than competitors.\" Lead with optimization.\n- Do NOT quote specific dollar amounts without running the formula in `references\u002Fcost-formula.md` first.\n- Do NOT recommend binary quantization without confirming model compatibility.\n- Do NOT dismiss the customer's cost concern. Acknowledge it, then optimize.\n- Do NOT include FDE ($200k) or Premium support in proposals unless the customer explicitly needs them.\n\n## Symptom 1: Customer says \"pricing is too expensive\"\n\n### When does this happen?\n- Customer states \"the price is killing us\" if they have to scale using RAM\n- Wants to reduce projected monthly cost from $500,000 to $50,000-$60,000\n- Current pricing proposal is too expensive for seed-stage startup\n\n### Diagnosis: What's driving the cost?\n\n**Step 1: Check what's in memory**\n- By default, Qdrant stores vectors on disk then memory-maps to keep\n  everything within memory. If someone explicitly sets `on_disk=false`\n  in vector configuration, that brings full resolution vectors into memory.\n- RAM is the number that drives the core unit cost. That's where most of\n  the operational and compute work happens with Qdrant.\n\n**Step 2: Check embedding dimensions**\n- Higher dimensions = more RAM = higher cost\n- The dimension model itself doesn't change Qdrant infrastructure pricing,\n  but it determines how much RAM each vector consumes\n\n**Step 3: Check quantization status**\n- \"Where are you at right now in the strategy around quantization and\n  trying to understand your real cutoff for accuracy as you compress\n  further? That right there is the single biggest driver of the RAM.\"\n\n### Solutions (ordered by impact)\n\n#### Option 1: Apply scalar quantization (4x RAM reduction)\n- **When to recommend:** Default first step. \"We recommend our customers\n  to use scalar if they want to try that first just to see what the\n  improvement is going to be in terms of cost.\"\n- **Best for:** \"90% of our customers that are using quantization are\n  usually using scalar. We very rarely offload vectors to disk just\n  because quantization works so well.\"\n- **Trade-off:** \"We're cutting our costs with quantization, but our\n  precision is lower than we would like.\"\n- **Mitigation:** Rescore against original vectors on disk. \"But now our\n  latency is 2 seconds instead of 1 second.\"\n- **Example:** \"You can go to product quantization with a 64x compression\n  ratio and then you would lower your cost. If I gave you a number of\n  like 42k, with product quantization you could go to like 36.\"\n- [Scalar quantization docs](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F#scalar-quantization)\n\n#### Option 2: Apply binary quantization (32x RAM reduction)\n- **When to recommend:** \"If you use OpenAI text embedding 3 small, it\n  was trained to be binary quantization compatible. Binary quantization\n  allows for a 32x reduction in the storage requirement. We also provide\n  1.5 and 2 bit quantization methods that are 16x and 24x reduction.\"\n- **Best for:** \"With rescoring we're able to go from like 70% typically\n  in terms of the recall to like 98% by rescoring the vectors using the\n  full resolution representation from disk.\"\n- **Trade-off:** \"There's a cost and that is the accuracy of your results.\n  That's something that you would have to test to be certain if that works.\"\n- **Example:** \"OpenAI text embeddings perform very well with binary\n  quantization with rescoring off disk. You fetch top 50 using binary\n  quantization, rank them, then grab the FP16 vectors from disk and\n  rescore them.\"\n- [Binary quantization docs](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F#binary-quantization)\n\n#### Option 3: Move vectors to disk (keep quantization in memory)\n- **When to recommend:** \"We tested, what about if we just persist the\n  vectors in disk and then have a small cache instead of having this\n  massive cache? The numbers were still a few seconds to retrieve.\"\n- **Best for:** \"The quantization is the one that is much needed, then\n  the HNSW and vector sectors can still be on the disk and we can still\n  pull them very fast.\"\n- **Trade-off:** \"What's slow is going to be the traversal through the\n  HNSW graph. Qdrant has to go to disk to fetch the links, the edges.\n  The fact that you have to hit disk maybe 200-250 times for a single\n  traversal means you're incurring cold start latencies for each fetch.\"\n- **Performance impact:** \"If you're traversing the HNSW graph in RAM,\n  it's nearly instantaneous, very low milliseconds. On disk, you're\n  entering high tens of milliseconds or possibly hundreds. It'll be\n  orders of magnitude faster by just keeping the graph in memory.\"\n- [Disk storage configuration](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fconcepts\u002Fstorage\u002F)\n\n#### Option 4: Reduce dimension size\n- **When to recommend:** \"At that scale, the loss of information from\n  binary and scalar quant in particular, you might actually have a bit\n  of benefit. You might not get much information loss because those\n  vectors are so huge.\"\n- **Example:** Use `text-embedding-3-small` truncated to 512 instead of\n  `text-embedding-3-large` at 3072.\n- [Embedding model selection](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F)\n\n#### Option 5: Store vectors as FP16 instead of FP32\n- **When to recommend:** \"You can truncate the resolution of the vectors\n  on disk to be float 16. That way the storage will be half of what\n  you're currently doing because you're probably using float 32.\"\n- **Trade-off:** \"Lose the precision, but especially if you're going to\n  be using the quantized vectors regardless, why would you not do it?\n  For retrieval tasks, we find it's often not even measurable in terms\n  of the impact on recall.\"\n- [Vector precision configuration](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fconcepts\u002Fvectors\u002F)\n\n## Symptom 2: Customer asks \"how much will it cost to scale?\"\n\n### When does this happen?\n- Customer needs predictable costing to pass to their client\n- Needs transparency: look at a meter, understand consumption, calculate\n  unit economics, create predictable forecasts\n- Needs a formula to calculate cost based on cluster size (RAM, CPU, disk)\n\n### Diagnosis: What's the scaling pattern?\n\n**Pattern 1: Linear scaling (same configuration)**\n- \"If you keep everything the same, keep the quantization method the same,\n  if you double the data set size, you're doubling your costs. It's just\n  like doubling of the storage and the RAM.\"\n\n**Pattern 2: Optimize at scale**\n- \"What we see is people don't necessarily fix a percentage of their data\n  set in RAM. They will simply keep the same nodes, same RAM config,\n  then add more data and put it all on disk until they say 'this is no\n  longer meeting my throughput SLAs and I need to increase RAM.'\"\n- \"When I do the calculation I typically try to make sure RAM utilization\n  is not above like an 80% threshold. That way there's still enough room\n  for the system to operate.\"\n\n### Solutions\n\n**Provide scaling scenarios:**\n- \"We have one option with the same quantization strategy and one with a\n  more aggressive strategy which we think can still hit your targets.\n  That's to be determined when we do some testing. This gives you a\n  range of where you might land.\"\n- Example: Show Year 1 (100M vectors, scalar) vs Year 2 (200M vectors, binary)\n- See `references\u002Fcost-formula.md` for the calculation methodology\n- [Capacity planning guide](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fcapacity-planning\u002F)\n\n## Symptom 3: Customer compares to competitor pricing\n\n### When does this happen?\n- Customer discovered pricing is 5x higher than competitors\n- Price difference may be due to quantization methods used by other vendors\n\n### Diagnosis: What are competitors doing?\n\n**Check competitor quantization:**\n- \"My experience with some competitors is that they'll often price people\n  for something like product and binary but then do a scalar in production,\n  and there's a discrepancy between the actual spend.\"\n- \"The RAM requirement for scalar would be 4x what you're seeing on the\n  left. You can imagine 4x the price essentially.\"\n\n### Solutions\n\n**Show apples-to-apples comparison:**\n- \"Scalar is the same compression ratio with RQ (rotational quantization).\n  What Weaviate does is a linear transformation on the vectors to rotate\n  them along an axis to make the variance more uniform, then scalar\n  quantize. We truncate to int8 but skip the preprocessing step because\n  we know the distribution from immutable segments.\"\n- Example: \"Competitor is $200k with binary quantization. We're $150k\n  with scalar quantization (better accuracy).\"\n- See `references\u002Fcompetitive-pricing.md` for specific comparisons\n\n## Symptom 4: Customer is seed-stage startup \u002F budget-constrained\n\n### When does this happen?\n- Startup says current pricing is too high for their stage\n- Customer did rough calculation, expected 3-5k, actual quote was 6x higher\n\n### Diagnosis: What's the minimum viable configuration?\n\n**Check current scale:**\n- What vector count? What growth rate?\n- \"If you looked into quantization, that's something we should definitely\n  check out to keep costs down.\"\n\n### Solutions\n\n**Start small, scale later:**\n- \"For 50 million vectors with scalar quantization: $600 a month. If you\n  just double it, you're looking at $1,300. For 100 million: $1,370.\"\n- \"With binary quantization with rescoring off disk, you're able to\n  achieve probably somewhere between 95-98% recall.\"\n- \"Let's say we went to binary quantization. That number drops down\n  close to $4,000 a month.\"\n\n## Common Patterns\n\n### \"Everything is on disk but cost is still high\"\n- **Root cause:** \"This is very CPU intensive software. Your RAM in that\n  case is like 9,000 gigabytes, so 9 terabytes. It's expensive.\"\n- **Fix:** Apply quantization first, then move to disk.\n\n### \"Quantization didn't reduce cost enough\"\n- **Root cause:** \"I just had a situation where I told somebody to set\n  quantized vectors to be in RAM and full resolution vectors on disk.\n  But they did sort of the opposite. People get tripped up on where\n  the configuration is.\"\n- **Fix:** Verify RAM configuration is set correctly. Otherwise you\n  don't benefit from the efficiency gains of quantization.\n\n### \"Need to reduce cost but can't sacrifice accuracy\"\n- **Root cause:** \"We would need to experiment with how important the\n  extra accuracy is. Could we get by with binary quantization and take\n  a little more noise to save cost, maybe decrease the latency because\n  2 seconds starts to become problematic. But if it was 500 milliseconds\n  with some decrease in accuracy, that might be where we want to end up.\"\n- **Fix:** Rescore off disk using full precision vectors. Traverse the\n  graph in memory using binary quantized vectors, grab top 100 results,\n  then fetch full precision vectors off disk for those 100 and re-rank.\n  Also oversample: if you want top 100, oversample by a factor of two.\n\n## Decision Tree\n\n```\nCustomer says \"too expensive\"\n  |\n  v\nWhat's driving cost?\n  |\n  +-- High RAM usage\n  |   |\n  |   +-- Full resolution vectors in memory?\n  |   |   -> Apply scalar quantization (4x reduction)\n  |   |   -> https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F#scalar-quantization\n  |   |\n  |   +-- Already using scalar?\n  |   |   -> Apply binary quantization (32x reduction)\n  |   |   -> https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F#binary-quantization\n  |   |\n  |   +-- Already using binary?\n  |       -> Move vectors to disk\n  |       -> https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fconcepts\u002Fstorage\u002F\n  |\n  +-- High dimension count\n  |   -> Reduce embedding dimensions\n  |   -> https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F\n  |\n  +-- Need predictable scaling\n      -> Provide scaling scenarios\n      -> See references\u002Fcost-formula.md\n```\n","# When and why to recommend cost reduction techniques\n\nThis skill converts SA knowledge into a navigation system over Qdrant\ndocumentation. It answers \"when?\" and \"why?\" for cost reduction, not \"how?\"\nEvery solution links to docs for implementation details.\n\n## Symptom routing\n\n| Customer signal | Go to | Load reference |\n|----------------|-------|----------------|\n| \"Too expensive\", sticker shock on proposal | Symptom 1 below | `references\u002Fcost-formula.md` |\n| \"How much will it cost at X scale?\" | Symptom 2 below | `references\u002Fcost-formula.md` |\n| \"Competitor is cheaper\" | Symptom 3 below | `references\u002Fcompetitive-pricing.md` |\n| Startup, budget-constrained | Symptom 4 below | `references\u002Fcost-formula.md` |\n\n## Do NOT recommend\n\n- Do NOT lead with \"Qdrant is cheaper than competitors.\" Lead with optimization.\n- Do NOT quote specific dollar amounts without running the formula in `references\u002Fcost-formula.md` first.\n- Do NOT recommend binary quantization without confirming model compatibility.\n- Do NOT dismiss the customer's cost concern. Acknowledge it, then optimize.\n- Do NOT include FDE ($200k) or Premium support in proposals unless the customer explicitly needs them.\n\n## Symptom 1: Customer says \"pricing is too expensive\"\n\n### When does this happen?\n- Customer states \"the price is killing us\" if they have to scale using RAM\n- Wants to reduce projected monthly cost from $500,000 to $50,000-$60,000\n- Current pricing proposal is too expensive for seed-stage startup\n\n### Diagnosis: What's driving the cost?\n\n**Step 1: Check what's in memory**\n- By default, Qdrant stores vectors on disk then memory-maps to keep\n  everything within memory. If someone explicitly sets `on_disk=false`\n  in vector configuration, that brings full resolution vectors into memory.\n- RAM is the number that drives the core unit cost. That's where most of\n  the operational and compute work happens with Qdrant.\n\n**Step 2: Check embedding dimensions**\n- Higher dimensions = more RAM = higher cost\n- The dimension model itself doesn't change Qdrant infrastructure pricing,\n  but it determines how much RAM each vector consumes\n\n**Step 3: Check quantization status**\n- \"Where are you at right now in the strategy around quantization and\n  trying to understand your real cutoff for accuracy as you compress\n  further? That right there is the single biggest driver of the RAM.\"\n\n### Solutions (ordered by impact)\n\n#### Option 1: Apply scalar quantization (4x RAM reduction)\n- **When to recommend:** Default first step. \"We recommend our customers\n  to use scalar if they want to try that first just to see what the\n  improvement is going to be in terms of cost.\"\n- **Best for:** \"90% of our customers that are using quantization are\n  usually using scalar. We very rarely offload vectors to disk just\n  because quantization works so well.\"\n- **Trade-off:** \"We're cutting our costs with quantization, but our\n  precision is lower than we would like.\"\n- **Mitigation:** Rescore against original vectors on disk. \"But now our\n  latency is 2 seconds instead of 1 second.\"\n- **Example:** \"You can go to product quantization with a 64x compression\n  ratio and then you would lower your cost. If I gave you a number of\n  like 42k, with product quantization you could go to like 36.\"\n- [Scalar quantization docs](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F#scalar-quantization)\n\n#### Option 2: Apply binary quantization (32x RAM reduction)\n- **When to recommend:** \"If you use OpenAI text embedding 3 small, it\n  was trained to be binary quantization compatible. Binary quantization\n  allows for a 32x reduction in the storage requirement. We also provide\n  1.5 and 2 bit quantization methods that are 16x and 24x reduction.\"\n- **Best for:** \"With rescoring we're able to go from like 70% typically\n  in terms of the recall to like 98% by rescoring the vectors using the\n  full resolution representation from disk.\"\n- **Trade-off:** \"There's a cost and that is the accuracy of your results.\n  That's something that you would have to test to be certain if that works.\"\n- **Example:** \"OpenAI text embeddings perform very well with binary\n  quantization with rescoring off disk. You fetch top 50 using binary\n  quantization, rank them, then grab the FP16 vectors from disk and\n  rescore them.\"\n- [Binary quantization docs](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F#binary-quantization)\n\n#### Option 3: Move vectors to disk (keep quantization in memory)\n- **When to recommend:** \"We tested, what about if we just persist the\n  vectors in disk and then have a small cache instead of having this\n  massive cache? The numbers were still a few seconds to retrieve.\"\n- **Best for:** \"The quantization is the one that is much needed, then\n  the HNSW and vector sectors can still be on the disk and we can still\n  pull them very fast.\"\n- **Trade-off:** \"What's slow is going to be the traversal through the\n  HNSW graph. Qdrant has to go to disk to fetch the links, the edges.\n  The fact that you have to hit disk maybe 200-250 times for a single\n  traversal means you're incurring cold start latencies for each fetch.\"\n- **Performance impact:** \"If you're traversing the HNSW graph in RAM,\n  it's nearly instantaneous, very low milliseconds. On disk, you're\n  entering high tens of milliseconds or possibly hundreds. It'll be\n  orders of magnitude faster by just keeping the graph in memory.\"\n- [Disk storage configuration](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fconcepts\u002Fstorage\u002F)\n\n#### Option 4: Reduce dimension size\n- **When to recommend:** \"At that scale, the loss of information from\n  binary and scalar quant in particular, you might actually have a bit\n  of benefit. You might not get much information loss because those\n  vectors are so huge.\"\n- **Example:** Use `text-embedding-3-small` truncated to 512 instead of\n  `text-embedding-3-large` at 3072.\n- [Embedding model selection](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F)\n\n#### Option 5: Store vectors as FP16 instead of FP32\n- **When to recommend:** \"You can truncate the resolution of the vectors\n  on disk to be float 16. That way the storage will be half of what\n  you're currently doing because you're probably using float 32.\"\n- **Trade-off:** \"Lose the precision, but especially if you're going to\n  be using the quantized vectors regardless, why would you not do it?\n  For retrieval tasks, we find it's often not even measurable in terms\n  of the impact on recall.\"\n- [Vector precision configuration](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fconcepts\u002Fvectors\u002F)\n\n## Symptom 2: Customer asks \"how much will it cost to scale?\"\n\n### When does this happen?\n- Customer needs predictable costing to pass to their client\n- Needs transparency: look at a meter, understand consumption, calculate\n  unit economics, create predictable forecasts\n- Needs a formula to calculate cost based on cluster size (RAM, CPU, disk)\n\n### Diagnosis: What's the scaling pattern?\n\n**Pattern 1: Linear scaling (same configuration)**\n- \"If you keep everything the same, keep the quantization method the same,\n  if you double the data set size, you're doubling your costs. It's just\n  like doubling of the storage and the RAM.\"\n\n**Pattern 2: Optimize at scale**\n- \"What we see is people don't necessarily fix a percentage of their data\n  set in RAM. They will simply keep the same nodes, same RAM config,\n  then add more data and put it all on disk until they say 'this is no\n  longer meeting my throughput SLAs and I need to increase RAM.'\"\n- \"When I do the calculation I typically try to make sure RAM utilization\n  is not above like an 80% threshold. That way there's still enough room\n  for the system to operate.\"\n\n### Solutions\n\n**Provide scaling scenarios:**\n- \"We have one option with the same quantization strategy and one with a\n  more aggressive strategy which we think can still hit your targets.\n  That's to be determined when we do some testing. This gives you a\n  range of where you might land.\"\n- Example: Show Year 1 (100M vectors, scalar) vs Year 2 (200M vectors, binary)\n- See `references\u002Fcost-formula.md` for the calculation methodology\n- [Capacity planning guide](https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fcapacity-planning\u002F)\n\n## Symptom 3: Customer compares to competitor pricing\n\n### When does this happen?\n- Customer discovered pricing is 5x higher than competitors\n- Price difference may be due to quantization methods used by other vendors\n\n### Diagnosis: What are competitors doing?\n\n**Check competitor quantization:**\n- \"My experience with some competitors is that they'll often price people\n  for something like product and binary but then do a scalar in production,\n  and there's a discrepancy between the actual spend.\"\n- \"The RAM requirement for scalar would be 4x what you're seeing on the\n  left. You can imagine 4x the price essentially.\"\n\n### Solutions\n\n**Show apples-to-apples comparison:**\n- \"Scalar is the same compression ratio with RQ (rotational quantization).\n  What Weaviate does is a linear transformation on the vectors to rotate\n  them along an axis to make the variance more uniform, then scalar\n  quantize. We truncate to int8 but skip the preprocessing step because\n  we know the distribution from immutable segments.\"\n- Example: \"Competitor is $200k with binary quantization. We're $150k\n  with scalar quantization (better accuracy).\"\n- See `references\u002Fcompetitive-pricing.md` for specific comparisons\n\n## Symptom 4: Customer is seed-stage startup \u002F budget-constrained\n\n### When does this happen?\n- Startup says current pricing is too high for their stage\n- Customer did rough calculation, expected 3-5k, actual quote was 6x higher\n\n### Diagnosis: What's the minimum viable configuration?\n\n**Check current scale:**\n- What vector count? What growth rate?\n- \"If you looked into quantization, that's something we should definitely\n  check out to keep costs down.\"\n\n### Solutions\n\n**Start small, scale later:**\n- \"For 50 million vectors with scalar quantization: $600 a month. If you\n  just double it, you're looking at $1,300. For 100 million: $1,370.\"\n- \"With binary quantization with rescoring off disk, you're able to\n  achieve probably somewhere between 95-98% recall.\"\n- \"Let's say we went to binary quantization. That number drops down\n  close to $4,000 a month.\"\n\n## Common Patterns\n\n### \"Everything is on disk but cost is still high\"\n- **Root cause:** \"This is very CPU intensive software. Your RAM in that\n  case is like 9,000 gigabytes, so 9 terabytes. It's expensive.\"\n- **Fix:** Apply quantization first, then move to disk.\n\n### \"Quantization didn't reduce cost enough\"\n- **Root cause:** \"I just had a situation where I told somebody to set\n  quantized vectors to be in RAM and full resolution vectors on disk.\n  But they did sort of the opposite. People get tripped up on where\n  the configuration is.\"\n- **Fix:** Verify RAM configuration is set correctly. Otherwise you\n  don't benefit from the efficiency gains of quantization.\n\n### \"Need to reduce cost but can't sacrifice accuracy\"\n- **Root cause:** \"We would need to experiment with how important the\n  extra accuracy is. Could we get by with binary quantization and take\n  a little more noise to save cost, maybe decrease the latency because\n  2 seconds starts to become problematic. But if it was 500 milliseconds\n  with some decrease in accuracy, that might be where we want to end up.\"\n- **Fix:** Rescore off disk using full precision vectors. Traverse the\n  graph in memory using binary quantized vectors, grab top 100 results,\n  then fetch full precision vectors off disk for those 100 and re-rank.\n  Also oversample: if you want top 100, oversample by a factor of two.\n\n## Decision Tree\n\n```\nCustomer says \"too expensive\"\n  |\n  v\nWhat's driving cost?\n  |\n  +-- High RAM usage\n  |   |\n  |   +-- Full resolution vectors in memory?\n  |   |   -> Apply scalar quantization (4x reduction)\n  |   |   -> https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F#scalar-quantization\n  |   |\n  |   +-- Already using scalar?\n  |   |   -> Apply binary quantization (32x reduction)\n  |   |   -> https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F#binary-quantization\n  |   |\n  |   +-- Already using binary?\n  |       -> Move vectors to disk\n  |       -> https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fconcepts\u002Fstorage\u002F\n  |\n  +-- High dimension count\n  |   -> Reduce embedding dimensions\n  |   -> https:\u002F\u002Fqdrant.tech\u002Fdocumentation\u002Fguides\u002Fquantization\u002F\n  |\n  +-- Need predictable scaling\n      -> Provide scaling scenarios\n      -> See references\u002Fcost-formula.md\n```\n",{"name":5,"license":15,"metadata":44,"description":7},{"tags":45,"author":50,"version":51,"category":52},[20,46,47,48,38,49],"pricing","quantization","deals","ram","thierry","1.0.0","sales","main","49e42b42210ce4eb860fa0745873674c57894331","qdrant-os\u002Fskills\u002Fqdrant-cost-reduction-advisor",1,"https:\u002F\u002Fgithub.com\u002Fqdrant-labs\u002Fdevrel-projects","https:\u002F\u002Fgithub.com\u002Fqdrant-labs","Experimental projects, community-driven tools, and works-in-progress built around Qdrant","2026-04-23 09:25:59",{"source":8,"skillName":5,"commands":62,"prompt":72,"zipAvailable":73},[63,66,69],{"tool":64,"command":65},"npx","npx skills add qdrant-labs\u002Fdevrel-projects --skill qdrant-cost-reduction-advisor",{"tool":67,"command":68},"bunx","bunx skills add qdrant-labs\u002Fdevrel-projects --skill qdrant-cost-reduction-advisor",{"tool":70,"command":71},"pnpm","pnpm dlx skills add qdrant-labs\u002Fdevrel-projects --skill qdrant-cost-reduction-advisor","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fqdrant-labs\u002Fdevrel-projects\u002Fblob\u002Fmain\u002Fqdrant-os\u002Fskills\u002Fqdrant-cost-reduction-advisor\u002FSKILL.md",true,[75,81],{"path":76,"resourceType":77,"resourceTypeDesc":78,"sizeBytes":79,"sourceHtmlUrl":80},"references\u002Fcost-formula.md","REFERENCE","参考文档",3153,"https:\u002F\u002Fgithub.com\u002Fqdrant-labs\u002Fdevrel-projects\u002Fblob\u002Fmain\u002Fqdrant-os\u002Fskills\u002Fqdrant-cost-reduction-advisor\u002Freferences\u002Fcost-formula.md",{"path":82,"resourceType":77,"resourceTypeDesc":78,"sizeBytes":83,"sourceHtmlUrl":84},"triggers.json",1450,"https:\u002F\u002Fgithub.com\u002Fqdrant-labs\u002Fdevrel-projects\u002Fblob\u002Fmain\u002Fqdrant-os\u002Fskills\u002Fqdrant-cost-reduction-advisor\u002Ftriggers.json",[86,114,144],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":97,"publishedAt":39},"01M71FAPNFM146C4F5AV36TDS4","qdrant-memory-optimization","Diagnoses and fixes Qdrant memory problems. Use when a developer reports OOM crashes, high RAM usage, HNSW consuming too much memory, latency spiking after switching storage, or needing to fit more vectors on existing hardware. Also trigger when someone asks which quantization to use, how to choose between mmap and RAM, how to estimate memory for scaling, or why their Qdrant node keeps dying. Do NOT use for initial Qdrant setup, API scripting, embedding model selection, search accuracy tuning, backups, or vector DB comparisons.","https:\u002F\u002Fgithub.com\u002Fqdrant-labs\u002Fdevrel-projects\u002Fblob\u002Fmain\u002Fqdrant-os\u002Fskills\u002Fqdrant-memory-optimization\u002FSKILL.md",[92],{"categoryNo":93,"categoryName":94,"taxonomyType":95,"taxonomyTypeDesc":96},"01JBSEEDCAT000000000000030","开发工程师","PROFESSION","适用职业",[21,98,99,100,101,102,103,104,38,105,106,107,108,109,110,35,111,112,34,113],"Optimization","Memory Optimization","qdrant memory optimization","memory","reports","which","embedding","model","db","choose","search","vector","selection","model selection","switching","api",{"skillNo":115,"name":116,"nameZh":6,"description":117,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":118,"categories":119,"keywords":128,"publishedAt":39},"01M71FAH49J72P540JWAVBRSKX","qdrant-internal-skill-creator","Guides Qdrant SAs through building internal skills that embed domain knowledge into Claude. Use when someone wants to \"create a qdrant skill\", \"build a skill for qdrant diagnostics\", \"make an internal skill for qdrant customers\", \"create a qdrant troubleshooting skill\", or asks how to \"turn SA knowledge into a skill\". Also trigger when someone says \"help me build a skill for qdrant deal objections\" or \"I want to create a skill for qdrant cluster sizing\". Do NOT use for general skill creation unrelated to Qdrant, external community skills, or Python\u002FJS project skills.","https:\u002F\u002Fgithub.com\u002Fqdrant-labs\u002Fdevrel-projects\u002Fblob\u002Fmain\u002Fqdrant-os\u002Fskills\u002Fqdrant-internal-skill-creator\u002FSKILL.md",[120,125],{"categoryNo":121,"categoryName":122,"taxonomyType":123,"taxonomyTypeDesc":124},"01JBSEEDCAT000000000000021","资料检索与研究","SCENARIO","使用场景",{"categoryNo":126,"categoryName":127,"taxonomyType":123,"taxonomyTypeDesc":124},"01JBSEEDCAT000000000000023","Skill与智能体开发",[129,130,21,131,24,132,133,134,29,135,136,137,138,139,35,140,34,141,142,143],"skill creator","qdrant internal skill creator","creator","make","diagnostics","a skill","unrelated","community","skill-creator","creation","project","knowledge","guides","create","me",{"skillNo":145,"name":146,"nameZh":6,"description":147,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":148,"categories":149,"keywords":151,"publishedAt":39},"01M71FA1G7KS1VX5SCBSJZXBA7","qdrant-community-skill-creator","Guides developers through building community Qdrant skills that embed domain knowledge into Claude. Use when someone wants to \"create a community qdrant skill\", \"build an external skill for qdrant\", \"make a qdrant skill for the community\", \"build a public qdrant skill\", or asks how to \"create a skill for qdrant memory issues for developers\". Also trigger when someone says \"I want to build an open source qdrant skill\" or \"help me create a qdrant skill for public use\". Do NOT use for internal SA skills, general non-Qdrant skill creation, or Python project skills unrelated to Qdrant.","https:\u002F\u002Fgithub.com\u002Fqdrant-labs\u002Fdevrel-projects\u002Fblob\u002Fmain\u002Fqdrant-os\u002Fskills\u002Fqdrant-community-skill-creator\u002FSKILL.md",[150],{"categoryNo":126,"categoryName":127,"taxonomyType":123,"taxonomyTypeDesc":124},[152,129,131,21,136,153,154,139,155,156,140,141,142,29,134,24,157,158,159,160],"qdrant community skill creator","python","python project","open","open source","domain","for","general","use"]